Feng Chen∗1,2, Botian Xu∗2, Pu Hua‡2,3,4, Peiqi Duan‡2, Yanchao Yang1, Yi Ma1, Huazhe Xu2,3,4
1The University of Hong Kong, 2Tsinghua University IIIS, 3Shanghai Qi Zhi Institute
4Shanghai AI Lab
cf24@connect.hku.hk,btx0424@outlook.com
huazhe_xu@mail.tsinghua.edu.cn
Abstract
Due to the difficulty of acquiring extensive real-world data, robot simulation has become crucial for parallel training and sim-to-real transfer, highlighting the importance of scalable simulated robotic tasks.Foundation models have demonstrated impressive capacities in autonomously generating feasible robotic tasks. However, this new paradigm underscores the challenge of adequately evaluating these autonomously generated tasks.To address this, we propose a comprehensive evaluation framework tailored to generative simulations.Our framework segments evaluation into three core aspects: quality, diversity, and generalization.For single-task quality, we evaluate the realism of the generated task and the completeness of the generated trajectories using large language models and vision-language models. In terms of diversity, we measure both task and data diversity through text similarity of task descriptions and world model loss trained on collected task trajectories. For task-level generalization, we assess the zero-shot generalization ability on unseen tasks of a policy trained with multiple generated tasks.Experiments conducted on three representative task generation pipelines demonstrate that the results from our framework are highly consistent with human evaluations, confirming the feasibility and validity of our approach.The findings reveal that while metrics of quality and diversity can be achieved through certain methods, no single approach excels across all metrics, suggesting a need for greater focus on balancing these different metrics. Additionally, our analysis further highlights the common challenge of low generalization capability faced by current works.Our anonymous website: https://sites.google.com/view/evaltasks.
1 Introduction
Embodied artificial intelligence(EAI) is crucial to enable intelligent agents to understand and interact with the physical world. However, creating such agents with physical forms and universal functionalities necessitates extensive data, which is prohibitively expensive to acquire manually(Dasari etal., 2020; Srivastava etal., 2021; Mu etal., 2021). Although multiple attempts have been made toward massive real-world data collection(Brohan etal., 2023b; a), training in simulated environments still plays a key role in various robotic tasks(Wang etal., 2023a; Huang etal., 2021; Lin etal., 2021; Yuan etal., 2024; Yu etal., 2020). Consequently, the acquisition of a substantial number of robotic tasks in simulation, which heavily rely on foundation models, is of significant importance.
Foundation models(OpenAI, 2023; Team etal., 2023; Zhang etal., 2023a) have exhibited remarkable proficiency in various robotics-related tasks, including coding(Rozière etal., 2023), 3D generation(Deitke etal., 2022; 2023), scene comprehension(Mohiuddin etal., 2024), planning(Huang etal., 2023b; 2024), and reward formulation(Ma etal., 2023). Notably, recent works have demonstrated the potential of leveraging such capabilities of foundation models to generate robotic tasks in simulation(Wang etal., 2023b; 2024; Katara etal., 2023; Yang etal., 2024; Hua etal., 2024).In generative simulation, foundation models such as large language models and vision-language models are prompted to output necessary task information(e.g., code, language descriptions), an appropriate scene, and successful trajectories for novel tasks at scale.However, despite these advancements, concerns have been raised regarding aspects such as the quality and reality of the generated tasks and whether the generated data can boost policy performance(Hua etal., 2024).Therefore, there is an urgent need for a comprehensive evaluation framework for generative simulation pipelines, which has so far been absent.
However, evaluating the generated tasks faces challenges similar to those encountered in assessing images (Salimans etal., 2016; Heusel etal., 2017) and texts (Wang etal., 2018), i.e., it is hard to quantify the realism of the generated tasks, thus hindering traditional evaluation mechanisms such as success rate from reflecting the quality and value of the generated tasks and data.In this paper, we propose a novel evaluation framework(see Fig 1) tailored to generative simulation pipelines.Our framework is concerned with three key perspectives: (1) the quality of single-task generation, which typically involves the alignment score of generated task scenarios with the real world and the completeness of generated task trajectories; (2) the task and data diversity concerned with generated tasks as well as generated trajectories; (3) the task-level generalization ability of a policy trained on a bunch of generated tasks.
Specifically, in terms of single-task quality, we leverage vision-language models to understand scene/trajectory images and output the scene alignment scores and task completion rates, which have been measured by human subjective judgment or hard-coded functions in previous works. Besides thorough single-task evaluation, we also incorporate multi-task evaluation on diversity and generalization, which are not extensively investigated in prior studies. For diversity, we first measure task diversity by examining the text similarity between the language descriptions of generated tasks. We then train a world model with trajectory data collected from these tasks and evaluate trajectory diversity based on the model’s prediction loss. For generalization, we train an imitation learning policy based on a variety of generated tasks and measure its efficacy on unseen tasks to gauge its task-level generalization capability.
In our experiments, we study 3 notable projects, namely, GenSim(Wang etal., 2024), RoboGen (Wang etal., 2023b), and BBSEA (Yang etal., 2024), in the hope of establishing a reference for subsequent research towards this direction. By comparing our evaluations with those from humans, we find that our evaluations reach consistent conclusions with the human experts on the vast majority of tasks.According to our evaluation results, RoboGen’s tasks exhibit the highest single-task quality while also holding an advantageous position in the textual task diversity of task descriptions. In terms of trajectory diversity, both GenSim and BBSEA have demonstrated superior results. Although currently none of the pipelines possess sufficiently excellent generalization capabilities, the tasks from GenSim still show a certain degree of potential in the direction of generalization.These findings indicate that although specific methods can achieve satisfactory quality and diversity metrics, none consistently outperform across all criteria, emphasizing the need for intensified efforts to balance these metrics effectively. Furthermore, our study also emphasizes the prevalent issue of low generalization capability encountered by current methodologies.
The main contributions of our work can be summarized as follows:
- •
We propose a novel framework to evaluate generative robotic simulation methods, providing researchers with tools to assess and improve their future works in this area.
- •
We develop an autonomous pipeline that quantitatively assesses the quality of an individual task using foundation models, which have previously been performed by human efforts.
- •
We introduce metrics for diversity and generalization of generated tasks and data to evaluate the value of multiple generated tasks and the extensive data derived from them.
2 Related Works
2.1 Foundation Models
In our article, we utilize large language models such as GPT-4 (OpenAI, 2023), released by OpenAI, which have had a profound impact on the field of natural language processing. Previous work has applied these large language models to the domain of robotics, specifically in policy learning (Driess etal., 2023; Huang etal., 2023c) and motion planning (Huang etal., 2022). Researchers have also explored using language models to generate code and rewards (Huang etal., 2023c; Wang etal., 2023b), aiding solvers in learning policies from tasks. Furthermore, vision-language models and multimodal foundational models have demonstrated remarkable potential (Zhang etal., 2023a; Team etal., 2023; Xu etal., 2023). Model GPT-4-vision (Zhang etal., 2023a) have exhibited capabilities in spatial understanding and basic assessment, making the automatic evaluation of tasks feasible. In prior work, vision-language models have been employed in the robotic task generation pipeline to verify the quality of the generated tasks (Wang etal., 2023b; 2024).
2.2 Generative Robotics Tasks and Datasets in Embodied AI
In recent research, foundational models have demonstrated remarkable capabilities (OpenAI, 2023; Zhang etal., 2023a; Team etal., 2023), leading to the emergence of autonomously generated robotic tasks in the field of robotics. Typically, such generative models utilize large language models to create a basic framework for generating tasks, which involves submitting the required 3D models through text-to-3D model (Li etal., 2023b) conversion, text-to-image (Mid, ) and image-to-3D models (Liu etal., 2023) processes, or searching and generating the necessary three-dimensional models in extensive 3D model repositories like Objaverse (Deitke etal., 2022; 2023). These models are then assembled into tasks within simulators, and methods such as reinforcement learning or trajectory optimization are employed to learn the trajectories needed to solve the tasks. Researchers have also explored tasks in other directions; for instance, the creators of Robogen have expanded task types to include soft materials and humanoid robots (Wang etal., 2023b), while the developers of Gensim have opted to deploy tasks on real robots, completing the generated tasks in the real world (Wang etal., 2024). Therefore, when evaluating the quality of generated tasks, it is also necessary to consider the diverse directions of exploration being pursued by different researchers.
2.3 Benchmarks on Meshes and Large Language Model
Recent work has bridged gaps in the evaluation of three-dimensional models and large language models. For instance, T3Bench (He etal., 2023) introduced the use of multiple foundational models to establish an evaluation system for metrics such as the quality and alignment of 3D models. The methods used in this work to assess the quality and alignment of 3D models have inspired our approach to evaluating the alignment of task scenarios in robotic tasks. Additionally, in the evaluation of large language models (Zhang etal., 2023b; Huang etal., 2023a), previous studies have discussed assessing various metrics across multiple scenarios to identify potential issues of hallucination and errors within models. These evaluation standards provide a valuable perspective for assessing robotic tasks, aiding in a more appropriate evaluation of such tasks.
2.4 Evaluation of Task Diversity
Learning a range of skills is crucial for building generalist robot agents that can autonomously operate in a complex environment. Therefore, we expect task generation to produce tasks with varying goals, dynamics, and environment setups such that collectively learning these tasks promotes generalization, robustness (Tobin etal., 2017), and even unseen skills (Eysenbach etal., 2018). However, evaluating such diversity of generated tasks remain unclear. RoboGen (Wang etal., 2023b) proposed to compare the Self-BLEU score and Embedding Similarity (Zhu etal., 2018) of the descriptions generated alongside the tasks. While such language-based diversity metrics consider high-level semantic information, they are strongly coupled with the language models used, which are known to be subject to alignment issues. In this work. we propose to evaluate task diversity as the coverage of skill or dynamics space, where high diversity facilitates better transfer or generalization to a held-out set of tasks. Recent model-based skill learning methods (Hafner etal., 2023; Hansen etal., 2024) are capable of learning highly complex dynamics and task information on a wide range of tasks. We leverage them for diversity evaluation.
3 Method
3.1 Introduction to Generative Simulation
Generative simulation represents a field of studies that utilize foundation models, particularly generative models pre-trained on internet-scale data, to acquire massive tasks and data in robot simulation.In generative simulation, large language models are first prompted to provide the basic framework for a novel task, such as the task description, assets, task code, etc. The task is then loaded into the simulation to construct a scene. We further query foundation models to provide objectives for task solutions, e.g. goals for planning or rewards for RL. Through RL training or motion planning, the pipeline will produce trajectory data for the previously generated task.To summarize, the performance of a generative simulation pipeline is fundamentally determined by key aspects such as the basic task framework, solution objectives, and the specific implementations of solution generation.
3.2 Overview
We divide our evaluation work into three parts, as visualized in Fig 2. In the first part(Sec 3.3), we assess the quality of a single generated task through foundational models, especially large language models and vision-language models, and statistical methods.In the second part(Sec 3.4), we respectively measure the diversity of generated task descriptions and trajectory data with a language model and a world model.In the third part(Sec 3.5), we evaluate the generalization capability of an imitation learning policy distilled from a large number of generated tasks.
3.3 Single-Task Quality
In this section, we introduce how we evaluate single-task quality. We consider two metrics: scene alignment score which measures the realism of the generated task and task completeness score which measures whether the generated task is successfully solved to collect data.
Scene alignment score. We utilize two different pipelines to evaluate scene alignment score.Due to possible deficiencies in visual recognition from foundation models(Tong etal., 2024b),one of our methods uses visual models, e.g., BLIP2(Li etal., 2023a), to generate textual descriptions of rendered scene images, followed by large language models(LLMs) such as GPT-4(OpenAI, 2023) to assess the consistency between the textual descriptions and the task descriptions. The other directly employs multi-modal LLMs like GPT-4V(Zhang etal., 2023a) and LLaVA(Liu etal., 2024) to evaluate the consistency between the task descriptions and the scene images.
Task completion score. We use foundation models, particularly vision-language models(VLMs), to assess the completion score of a generated task trajectory. Previously, this assessment was conducted using hard-coded functions, which demonstrated only limited capability in measuring task completion for specific tasks.Specifically, our approach begins with generating a video of the robot’s trajectory as we execute the task solution. From the video, we extract 8 images and provide them, along with the task description, to a VLM. We then obtain an evaluation of the task completion status from the VLM.
To reduce possible inherent biases and instability within foundation models, we conduct multiple scoring iterations and take the mean scores when evaluating on both metrics.
3.4 Task and Data Diversity
The generated tasks are expected to be diverse so that training on these tasks grants agents a range of skills and the ability to operate in various situations. However, a concrete definition of diversity is hard: in what sense are tasks distinct or similar? In this work, we are concerned with diversity from the following perspectives: (1) task diversity, a high-level diversity as identified by LLMs; and (2) trajectory diversity, a low-level diversity in terms of the dynamics of the collected data.
Text-based task diversity. Since LLMs generate tasks including the task descriptions and possibly scene configurations and goals, they are supposed to have an internal understanding of diversity at a high level. For example, “stack-blocks-tower” differs from “align-balls” semantically in terms of the action (verb) and the object of interest. Therefore, the similarity between embeddings of task descriptions can be considered as the similarity between tasks. Specifically, following (Zhu etal., 2018), we compute the diversity of a task set with text embeddings as:
(1) |
where is the number of tasks and removes self-similarity. A higher value indicates lower similarity and hence higher diversity.
Dynamics-based trajectory diversity. Though straightforward, task description diversity itself does not sufficiently characterize the actual learning experience of tasks, e.g., different interaction dynamics will take place when training on different generated tasks. Ideally, a diverse set of tasks should cover a large space of dynamics to promote the agent’s robustness under different scenarios. Therefore, we propose to evaluate such diversity through prediction error of dynamics models. Dynamics prediction error has been associated with novelty and widely adopted to promote exploration (Pathak etal., 2017; Burda etal., 2018). A high dynamics prediction error indicates unfamiliar (and thus novel) dynamics being experienced. We leverage a latent dynamics model following DreamerV3 (Hafner etal., 2024), where and are the observation and action at time step . The model is trained on trajectories collected from the generated tasks and then evaluated to compute the prediction errors.As will be discussed in Section 4.2, this approach helps us to identify tasks that render notably similar dynamics and are therefore not diverse.
3.5 Task Generalization
Generalization can be an ambiguous yet vital metric for evaluating the capabilities of generalist robot agents.In this paper, we define generalization as the capability to solve tasks within the same distribution, specifically whether an agent trained on the generated tasks can address similar scenarios and objectives albeit with varying initial states and minor low-level variations. To quantitatively examine this capability, we first train an imitation learning policy with trajectories collected by either the oracle policies or policies learned from the generation pipeline. The trained policy is subsequently evaluated with new task scenarios including varied object instances, appearance, and initial poses. The policy uses the state-of-the-art algorithm called Diffusion Policy (Chi etal., 2023) as the backbone and takes as input RGB observations, and the proprioceptions. A typical indicator of low generalization is when the trained policy performs well on the training data, confirming correct algorithm implementation, yet struggles to adapt to the varied tasks during evaluation.
4 Experiment
4.1 Single Task Evaluation
Experimental setup. As mentioned in Section3.3, our methodology utilizes vision-language models(VLMs) to generate scene captions, which are then compared against task descriptions using large language models(LLMs). Additionally, we employ a multi-modal LLM(MLLM) to evaluate the completeness of task trajectories. For captioning scene images, we deploy several VLMs, including BLIP (“blip2-flan-t5-xl”), Cambrian(“Cambrian-8B”)(Tong etal., 2024a), and LLaVA 1.6 (“LLaVA-1.6-7B”)(Liu etal., 2024). The scene alignment score is measured using the GPT-4(“2024-02-15-preview” version from Microsoft Azure) model. For assessing task completion, the MLLM models used include GPT-4V, Cambrian, and LLaVA.
4.1.1 Human verification
To validate the efficacy of our method, we gather human evaluations for ten tasks from the released tasks of RoboGen and GenSim and examine the consistency of our results with human results. We characterize their relationship by using the Pearson correlation coefficient to represent correlation strength and the mean absolute error (MAE) to indicate numerical similarity. A higher Pearson correlation coefficient signifies a stronger correlation, while a lower MAE reflects greater similarity. Therefore, we calculate the ratio of the Pearson correlation coefficient to the MAE to assess the relationship between our method and human evaluations; higher values indicate greater similarity.
In Figure3 left, in terms of scene alignment score, the performance of architectures using GPT-4 and other vision models like BLIP2 and LLaVA is shown to yield better results for RoboGen’s tasks, but these models perform poorly on GenSim’s tasks, primarily due to their lack of knowledge regarding top-view rendered images. In addition, Figure3 right illustrates that for task completion score, GPT-4V exhibits relative performance compared to human evaluations, indicating a strong alignment with human behavior in assessing task completion. In contrast, Cambrian and LLaVA 1.6 produce results that do not correspond with human assessments. While both models have demonstrated an understanding of images during the experiments, they fail to provide completion scores that align with human evaluations based on the image results.
4.1.2 Evaluation on RoboGen, GenSim, and BBSEA
We summarize the evaluation results of both metrics for single-task quality in Figure4.To be specific, among the methods, RoboGen tasks demonstrate high task completion scores, but their scene alignment scores are notably low. This discrepancy arises because, although RoboGen can generate assets relevant to the current task, these assets often collide when loaded into the scene, resulting in a cluttered and difficult-to-recognize environment. In contrast, GenSim secures the highest scene alignment scores but underperforms in task completion. This shortfall is largely attributed to its vision-language model lacking access to top-view rendered data, which impairs its ability to accurately recognize task completion. In addition, BBSEA achieves decent results on both metrics (although not the best), and it has the smallest variance in the outcomes.
Furthermore, we observe performance discrepancies between published and newly generated tasks across all methods. While a predictable decline in performance for generated tasks can be attributed to additional filtering prior to project release, improvements have been noted in GenSim’s scene alignment and task completion for RoboGen and BBSEA. The underlying reason is the advancements in the performance of foundation models, which have expanded the limits of task generation quality, including reasonable solution objectives in RoboGen and BBSEA, and innovative long-horizon task proposals in GenSim.
4.1.3 Examples from Evaluation
As shown in Figure5, in the snapshot of the ‘Open Laptop’ task, the laptop is correctly placed on the table, and there are some objects such as a lamp and a pen placed on the desk. Then we can observe from three trajectory images that the robotic arm has correctly located the laptop and opened it. Therefore, this task gets an average score of ‘7.96’ (out of 10) for completion score and an average score of ‘3.96’ (out of 5) for scene alignment.In the snapshot of the ‘Connect boxes with rope’, although we can abstract the red balls into a rope, there is also a gap between the scene and the real world, which gets ‘2.80’ in scene alignment. But when the red balls are expanded into a line, our pipeline can correctly figure out the task has been solved, thus receiving a ‘6.00’ completion score.In the snapshot of the ‘Move the stick into the green bin’, the gripper has grasped the stick and put it on the green bin rather than into the green bin, therefore our pipeline grades the task completion with ‘2.40’. But the scene receives ‘4.00’ because there are necessary objects, as well as some relevant ones on the table in the scene.
4.2 Task and Data Diversity
In this section, we use the proposed evaluation protocols to examine whether a pipeline generates diverse tasks and hence diverse data for learning. To have a better perspective for analysis and allow practical training, we divide the tasks into groups according to skills, scene configuration, or objects involved. The details for grouping can be found in SectionA.3.
Task diversity. For text-based task diversity, we leverage different language models, including “MiniLM-L6-v2” and “Mpnet-base-v2” from SentenceTransformers (Reimers & Gurevych, 2019; 2020), “LLama-3.1-70B” and “LLama-3.2-90B” (Touvron etal., 2023; Dubey etal., 2024), to generate text embeddings from task descriptions. The diversity is then measured by embedding similarity between and among the template and generated tasks following Equation1. We consider different task groups as well as the whole task set for each method. The results are listed in Table1, with higher values indicating higher diversity.Among the three methods, BBSEA shows the highest task description diversity by our proposed metric(1). However, we observe that many drawer tasks share remarkably similar descriptions, e.g., “open the drawer using handle”. Accordingly, the diversity of table-top tasks is significantly higher than that of drawer tasks.Conversely, results in GenSim indicate low task diversity because GenSim only deals with table-top pick-place tasks, narrowing its task domain.In addition, despite the extensive task types and complicated scenes RoboGen can support, it acquires lower scores than BBSEA. We attribute this underperformance to some vague task descriptions generated by RoboGen, which hinders the text similarity from reflecting task diversity.
Moreover, for all methods, we observe notable inconsistency between the language models from which we obtain the embeddings, possibly due to the difference in training method and objective that make the models attend to different components of the descriptions. This suggests that despite simplicity, text evaluation does not consistently and reliably capture the diversity of generated tasks.
Method | Task Group | MiniLM-L6-v2 | Mpnet-base-v2 | LLama-3.1-70B | LLama-3.2-90B |
GenSim | Stacking | 0.49 | 0.48 | 0.23 | 0.33 |
Placement | 0.55 | 0.52 | 0.12 | 0.15 | |
Piles | 0.42 | 0.45 | 0.26 | 0.40 | |
Assembling | 0.53 | 0.44 | 0.20 | 0.29 | |
All | 0.75 | 0.70 | 0.25 | 0.34 | |
RoboGen | Table-Top | 0.88 | 0.71 | 0.22 | 0.26 |
Ground | 0.78 | 0.65 | 0.24 | 0.24 | |
All | 0.84 | 0.69 | 0.25 | 0.28 | |
BBSEA | Table-Top | 1.22 | 1.06 | 0.24 | 0.26 |
Drawer | 0.37 | 0.34 | 0.27 | 0.28 | |
All | 1.28 | 1.10 | 0.28 | 0.29 |
Trajectory diversity. In terms of dynamics-based trajectory diversity, we examine whether the generated tasks provide trajectories with diverse dynamics coverage with a world model. Specifically, a total of 40 episodes is collected for each task in a group using the policy for each method. We train a world model in DreamerV3 using 10, 20, and 40 episodes respectively, and evaluate all 40 episodes.Please refer to AppendixA for details.Intuitively, a task group with low diversity is likely to exhibit comparably low prediction errors across various numbers of episodes, as a small dataset is sufficient to capture the dynamics. Conversely, for a more diverse task group, the prediction error of the world model should decrease as the volume of training data increases.
Method | Task Group | Eval error on 10 ep | Eval error on 20 ep | Eval error on 40 ep |
GenSim | Stacking | 115.0 | 67.2 | 24.5 |
Placement | 245.0 | 177.6 | 29.5 | |
Piles | 68.3 | 47.6 | 14.5 | |
Assembling | 105.4 | 64.3 | 29.6 | |
RoboGen | Table-Top | 16.2 | 10.9 | 6.4 |
Ground | 45.8 | 28.6 | 14.5 | |
BBSEA | Drawer | 402.4 | 287.0 | 69.6 |
Table-Top | 561.6 | 296.6 | 87.9 |
The results are shown in Table2. For GenSim, the model evaluation error for group Piles (where the objects of interest are piles of pellets) is significantly lower than others. This aligns with the fact that all tasks in this group only involve pushing piles on the table to a specified location. On the other hand, Placement, involves placing different types of objects in different manners, showing a much model higher error when trained on a small number of trajectories.Regarding BBSEA, cases are similar to those in GenSim.However, for RoboGen, both groups exhibit low model errors. This is primarily due to RoboGen’s learning design: for each task, all trajectories start with the same initial state. Therefore, the dynamics seen by the agent are similar and easy to learn by the world model.
4.3 Generalization
MethodGenSimRoboGenBBSEATask GroupStackingPlacementPilesAssemblingTable-TopGroundTable-TopDrawerReward/Success Rate0.150.320.20.080.00.00.000.00
We train a Diffusion Policy for each task group using 40 trajectories, identical to the data for dynamics model training, and assess their performance on the same task group under variations such as scene configurations, object colors, and initial robot states. As indicated in Table3, GenSim demonstrates reasonable generalization performance, despite the challenge posed by randomizing object colors, which complicates the effectiveness of an RGB-based policy. Conversely, agents trained on RoboGen and BBSEA tasks exhibit poor generalization. For RoboGen, the primary issue is that the training trajectories all begin from the same initial state, and the RL solutions do not generate high-quality data. In the case of BBSEA, the problem often lies in the repetition of similar tasks, which restricts task-level generalization capabilities. Moreover, significant task variations can result in out-of-distribution challenges that adversely affect agent performance.
5 Conclusion and Discussion
In this paper, we propose a novel evaluation framework for generative simulation, which includes three fundamental metrics: quality, diversity and generalization.We evaluate three representative generative simulation pipelines based on our proposed method.Results indicate that while various pipelines excel in terms of quality and diversity, there remains significant potential for improvement in their generalization capabilities. We hope that future work in generative simulation can make advancements and improvements in these three areas, especially in terms of generalization.
Moreover, we identify and outline some common drawbacks and failure cases across current generative simulation pipelines as follows for instructions to encourage further exploration:
- •
Low-quality task descriptions: Although task proposal is not a bottleneck for generative simulation in general, we still observe some vague and repeated task descriptions that fail to express the details of the generated tasks. Such ambiguity may cause suboptimal results in the evaluation of text-based task diversity, as well as harm the performance of a language-conditioned policy.
- •
Trajectory data with limited diversity: The task solution in some methods only considers limited task and scene variations, which will affect both the trajectory diversity and task-level generalization capability. Typical cases include insufficient intra-task randomization, relatively fixed task domain, or identical semantics between different tasks, leading to very similar trajectories. We advocate an appropriate dynamics model trained along with the task generation process to inspect and improve the diversity regarding dynamics coverage in future works.
- •
Task-specific design data collection and imitation learning: We remark that, on designing the task generation pipeline, generalization could be considered and improved in various ways, e.g., action space with good abstraction (control by end-effector poses, joint positions, or primitive actions), data augmentation, and unified goal specification. For example, in GenSim, actions are abstracted as high-level waypoints, and each task trajectory contains only a few such high-level actions. This design benefits its generalization evaluation based on imitation learning. We aim to devise a generally applicable protocol, with the diffusion policy not tuned or adopted specifically, and advocate efficient domain-specific designs for better generalization performance.
Acknowledgement
We want to thank Zhecheng Yuan, and Sizhe Yang for their generous help in implementing GenSim and BBSEA. We thank Brent Yi, Yuexiang Zhai for the discussion.
References
- (1)Midjourney.https://www.midjourney.com/home/.
- Brohan etal. (2023a)Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, XiChen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, Pete Florence, Chuyuan Fu, MontseGonzalez Arenas, Keerthana Gopalakrishnan, Kehang Han, Karol Hausman, Alexander Herzog, Jasmine Hsu, Brian Ichter, Alex Irpan, Nikhil Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Isabel Leal, Lisa Lee, Tsang-WeiEdward Lee, Sergey Levine, Yao Lu, Henryk Michalewski, Igor Mordatch, Karl Pertsch, Kanishka Rao, Krista Reymann, Michael Ryoo, Grecia Salazar, Pannag Sanketi, Pierre Sermanet, Jaspiar Singh, Anikait Singh, Radu Soricut, Huong Tran, Vincent Vanhoucke, Quan Vuong, Ayzaan Wahid, Stefan Welker, Paul Wohlhart, Jialin Wu, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Tianhe Yu, and Brianna Zitkovich.Rt-2: Vision-language-action models transfer web knowledge to robotic control, 2023a.
- Brohan etal. (2023b)Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, Julian Ibarz, Brian Ichter, Alex Irpan, Tomas Jackson, Sally Jesmonth, NikhilJ Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Isabel Leal, Kuang-Huei Lee, Sergey Levine, Yao Lu, Utsav Malla, Deeksha Manjunath, Igor Mordatch, Ofir Nachum, Carolina Parada, Jodilyn Peralta, Emily Perez, Karl Pertsch, Jornell Quiambao, Kanishka Rao, Michael Ryoo, Grecia Salazar, Pannag Sanketi, Kevin Sayed, Jaspiar Singh, Sumedh Sontakke, Austin Stone, Clayton Tan, Huong Tran, Vincent Vanhoucke, Steve Vega, Quan Vuong, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Tianhe Yu, and Brianna Zitkovich.Rt-1: Robotics transformer for real-world control at scale, 2023b.
- Burda etal. (2018)Yuri Burda, Harrison Edwards, Amos Storkey, and Oleg Klimov.Exploration by random network distillation.arXiv preprint arXiv:1810.12894, 2018.
- Chi etal. (2023)Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song.Diffusion policy: Visuomotor policy learning via action diffusion.In Proceedings of Robotics: Science and Systems (RSS), 2023.
- Dasari etal. (2020)Sudeep Dasari, Frederik Ebert, Stephen Tian, Suraj Nair, Bernadette Bucher, Karl Schmeckpeper, Siddharth Singh, Sergey Levine, and Chelsea Finn.Robonet: Large-scale multi-robot learning, 2020.
- Deitke etal. (2022)Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi.Objaverse: A universe of annotated 3d objects, 2022.
- Deitke etal. (2023)Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, SamirYitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi.Objaverse-xl: A universe of 10m+ 3d objects, 2023.
- Driess etal. (2023)Danny Driess, Fei Xia, Mehdi S.M. Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, Wenlong Huang, Yevgen Chebotar, Pierre Sermanet, Daniel Duckworth, Sergey Levine, Vincent Vanhoucke, Karol Hausman, Marc Toussaint, Klaus Greff, Andy Zeng, Igor Mordatch, and Pete Florence.Palm-e: An embodied multimodal language model, 2023.
- Dubey etal. (2024)Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, Arun Rao, Aston Zhang, Aurelien Rodriguez, Austen Gregerson, Ava Spataru, Baptiste Roziere, Bethany Biron, Binh Tang, Bobbie Chern, Charlotte Caucheteux, Chaya Nayak, Chloe Bi, Chris Marra, Chris McConnell, Christian Keller, Christophe Touret, Chunyang Wu, Corinne Wong, CristianCanton Ferrer, Cyrus Nikolaidis, Damien Allonsius, Daniel Song, Danielle Pintz, Danny Livshits, David Esiobu, Dhruv Choudhary, Dhruv Mahajan, Diego Garcia-Olano, Diego Perino, Dieuwke Hupkes, Egor Lakomkin, Ehab AlBadawy, Elina Lobanova, Emily Dinan, EricMichael Smith, Filip Radenovic, Frank Zhang, Gabriel Synnaeve, Gabrielle Lee, GeorgiaLewis Anderson, Graeme Nail, Gregoire Mialon, Guan Pang, Guillem Cucurell, Hailey Nguyen, Hannah Korevaar, HuXu, Hugo Touvron, Iliyan Zarov,ImanolArrieta Ibarra, Isabel Kloumann, Ishan Misra, Ivan Evtimov, Jade Copet, Jaewon Lee, Jan Geffert, Jana Vranes, Jason Park, Jay Mahadeokar, Jeet Shah, Jelmer vander Linde, Jennifer Billock, Jenny Hong, Jenya Lee, Jeremy Fu, Jianfeng Chi, Jianyu Huang, Jiawen Liu, Jie Wang, Jiecao Yu, Joanna Bitton, Joe Spisak, Jongsoo Park, Joseph Rocca, Joshua Johnstun, Joshua Saxe, Junteng Jia, KalyanVasuden Alwala, Kartikeya Upasani, Kate Plawiak, KeLi, Kenneth Heafield, Kevin Stone, Khalid El-Arini, Krithika Iyer, Kshitiz Malik, Kuenley Chiu, Kunal Bhalla, Lauren Rantala-Yeary, Laurens vander Maaten, Lawrence Chen, Liang Tan, Liz Jenkins, Louis Martin, Lovish Madaan, Lubo Malo, Lukas Blecher, Lukas Landzaat, Luke deOliveira, Madeline Muzzi, Mahesh Pasupuleti, Mannat Singh, Manohar Paluri, Marcin Kardas, Mathew Oldham, Mathieu Rita, Maya Pavlova, Melanie Kambadur, Mike Lewis, Min Si, MiteshKumar Singh, Mona Hassan, Naman Goyal, Narjes Torabi, Nikolay Bashlykov, Nikolay Bogoychev, Niladri Chatterji, OlivierDuchenne, Onur Çelebi, Patrick Alrassy, Pengchuan Zhang, Pengwei Li, Petar Vasic, Peter Weng, Prajjwal Bhargava, Pratik Dubal, Praveen Krishnan, PunitSingh Koura, Puxin Xu, Qing He, Qingxiao Dong, Ragavan Srinivasan, Raj Ganapathy, Ramon Calderer, RicardoSilveira Cabral, Robert Stojnic, Roberta Raileanu, Rohit Girdhar, Rohit Patel, Romain Sauvestre, Ronnie Polidoro, Roshan Sumbaly, Ross Taylor, Ruan Silva, Rui Hou, Rui Wang, Saghar Hosseini, Sahana Chennabasappa, Sanjay Singh, Sean Bell, SeohyunSonia Kim, Sergey Edunov, Shaoliang Nie, Sharan Narang, Sharath Raparthy, Sheng Shen, Shengye Wan, Shruti Bhosale, Shun Zhang, Simon Vandenhende, Soumya Batra, Spencer Whitman, Sten Sootla, Stephane Collot, Suchin Gururangan, Sydney Borodinsky, Tamar Herman, Tara Fowler, Tarek Sheasha, Thomas Georgiou, Thomas Scialom, Tobias Speckbacher, Todor Mihaylov, Tong Xiao, Ujjwal Karn, Vedanuj Goswami, Vibhor Gupta, Vignesh Ramanathan, Viktor Kerkez, Vincent Gonguet, Virginie Do, Vish Vogeti, Vladan Petrovic, Weiwei Chu,Wenhan Xiong, Wenyin Fu, Whitney Meers, Xavier Martinet, Xiaodong Wang, XiaoqingEllen Tan, Xinfeng Xie, Xuchao Jia, Xuewei Wang, Yaelle Goldschlag, Yashesh Gaur, Yasmine Babaei, YiWen, Yiwen Song, Yuchen Zhang, Yue Li, Yuning Mao, ZacharieDelpierre Coudert, Zheng Yan, Zhengxing Chen, Zoe Papakipos, Aaditya Singh, Aaron Grattafiori, Abha Jain, Adam Kelsey, Adam Shajnfeld, Adithya Gangidi, Adolfo Victoria, Ahuva Goldstand, Ajay Menon, Ajay Sharma, Alex Boesenberg, Alex Vaughan, Alexei Baevski, Allie Feinstein, Amanda Kallet, Amit Sangani, Anam Yunus, Andrei Lupu, Andres Alvarado, Andrew Caples, Andrew Gu, Andrew Ho, Andrew Poulton, Andrew Ryan, Ankit Ramchandani, Annie Franco, Aparajita Saraf, Arkabandhu Chowdhury, Ashley Gabriel, Ashwin Bharambe, Assaf Eisenman, Azadeh Yazdan, Beau James, Ben Maurer, Benjamin Leonhardi, Bernie Huang, Beth Loyd, BetoDe Paola, Bhargavi Paranjape, Bing Liu, BoWu, Boyu Ni, Braden Hancock, Bram Wasti, Brandon Spence, Brani Stojkovic, Brian Gamido, Britt Montalvo, CarlParker, Carly Burton, Catalina Mejia, Changhan Wang, Changkyu Kim, Chao Zhou, Chester Hu, Ching-Hsiang Chu, Chris Cai, Chris Tindal, Christoph Feichtenhofer, Damon Civin, Dana Beaty, Daniel Kreymer, Daniel Li, Danny Wyatt, David Adkins, David Xu, Davide Testuggine, Delia David, Devi Parikh, Diana Liskovich, Didem Foss, Dingkang Wang, Duc Le, Dustin Holland, Edward Dowling, Eissa Jamil, Elaine Montgomery, Eleonora Presani, Emily Hahn, Emily Wood, Erik Brinkman, Esteban Arcaute, Evan Dunbar, Evan Smothers, Fei Sun, Felix Kreuk, Feng Tian, Firat Ozgenel, Francesco Caggioni, Francisco Guzmán, Frank Kanayet, Frank Seide, GabrielaMedina Florez, Gabriella Schwarz, Gada Badeer, Georgia Swee, Gil Halpern, Govind Thattai, Grant Herman, Grigory Sizov, Guangyi, Zhang, Guna Lakshminarayanan, Hamid Shojanazeri, Han Zou, Hannah Wang, Hanwen Zha, Haroun Habeeb, Harrison Rudolph, Helen Suk, Henry Aspegren, Hunter Goldman, Ibrahim Damlaj, Igor Molybog, Igor Tufanov, Irina-Elena Veliche, Itai Gat, Jake Weissman, JamesGeboski, James Kohli, Japhet Asher, Jean-Baptiste Gaya, Jeff Marcus, Jeff Tang, Jennifer Chan, Jenny Zhen, Jeremy Reizenstein, Jeremy Teboul, Jessica Zhong, Jian Jin, Jingyi Yang, Joe Cummings, Jon Carvill, Jon Shepard, Jonathan McPhie, Jonathan Torres, Josh Ginsburg, Junjie Wang, Kai Wu, KamHou U, Karan Saxena, Karthik Prasad, Kartikay Khandelwal, Katayoun Zand, Kathy Matosich, Kaushik Veeraraghavan, Kelly Michelena, Keqian Li, Kun Huang, Kunal Chawla, Kushal Lakhotia, Kyle Huang, Lailin Chen, Lakshya Garg, Lavender A, Leandro Silva, Lee Bell, Lei Zhang, Liangpeng Guo, Licheng Yu, Liron Moshkovich, Luca Wehrstedt, Madian Khabsa, Manav Avalani, Manish Bhatt, Maria Tsimpoukelli, Martynas Mankus, Matan Hasson, Matthew Lennie, Matthias Reso, Maxim Groshev, Maxim Naumov, Maya Lathi, Meghan Keneally, MichaelL. Seltzer, Michal Valko, Michelle Restrepo, Mihir Patel, Mik Vyatskov, Mikayel Samvelyan, Mike Clark, Mike Macey, Mike Wang, MiquelJubert Hermoso, MoMetanat, Mohammad Rastegari, Munish Bansal, NandhiniSanthanam, Natascha Parks, Natasha White, Navyata Bawa, Nayan Singhal, Nick Egebo, Nicolas Usunier, NikolayPavlovich Laptev, Ning Dong, Ning Zhang, Norman Cheng, Oleg Chernoguz, Olivia Hart, Omkar Salpekar, Ozlem Kalinli, Parkin Kent, Parth Parekh, Paul Saab, Pavan Balaji, Pedro Rittner, Philip Bontrager, Pierre Roux, Piotr Dollar, Polina Zvyagina, Prashant Ratanchandani, Pritish Yuvraj, Qian Liang, Rachad Alao, Rachel Rodriguez, Rafi Ayub, Raghotham Murthy, Raghu Nayani, Rahul Mitra, Raymond Li, Rebekkah Hogan, Robin Battey, Rocky Wang, Rohan Maheswari, Russ Howes, Ruty Rinott, SaiJayesh Bondu, Samyak Datta, Sara Chugh, Sara Hunt, Sargun Dhillon, Sasha Sidorov, Satadru Pan, Saurabh Verma, Seiji Yamamoto, Sharadh Ramaswamy, Shaun Lindsay, Shaun Lindsay, Sheng Feng, Shenghao Lin, ShengxinCindy Zha, Shiva Shankar, Shuqiang Zhang, Shuqiang Zhang, Sinong Wang, Sneha Agarwal, Soji Sajuyigbe, Soumith Chintala, Stephanie Max, Stephen Chen, Steve Kehoe, Steve Satterfield, Sudarshan Govindaprasad, Sumit Gupta,Sungmin Cho, Sunny Virk, Suraj Subramanian, SyChoudhury, Sydney Goldman, Tal Remez, Tamar Glaser, Tamara Best, Thilo Kohler, Thomas Robinson, Tianhe Li, Tianjun Zhang, Tim Matthews, Timothy Chou, Tzook Shaked, Varun Vontimitta, Victoria Ajayi, Victoria Montanez, Vijai Mohan, VinaySatish Kumar, Vishal Mangla, Vítor Albiero, Vlad Ionescu, Vlad Poenaru, VladTiberiu Mihailescu, Vladimir Ivanov, Wei Li, Wenchen Wang, Wenwen Jiang, Wes Bouaziz, Will Constable, Xiaocheng Tang, Xiaofang Wang, Xiaojian Wu, Xiaolan Wang, Xide Xia, Xilun Wu, Xinbo Gao, Yanjun Chen, YeHu, YeJia, YeQi, Yenda Li, Yilin Zhang, Ying Zhang, Yossi Adi, Youngjin Nam, Yu, Wang, Yuchen Hao, Yundi Qian, Yuzi He, Zach Rait, Zachary DeVito, Zef Rosnbrick, Zhaoduo Wen, Zhenyu Yang, and Zhiwei Zhao.The llama 3 herd of models, 2024.URL https://arxiv.org/abs/2407.21783.
- Eysenbach etal. (2018)Benjamin Eysenbach, Abhishek Gupta, Julian Ibarz, and Sergey Levine.Diversity is all you need: Learning skills without a reward function.ArXiv, abs/1802.06070, 2018.URL https://api.semanticscholar.org/CorpusID:3521071.
- Hafner etal. (2023)Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy Lillicrap.Mastering diverse domains through world models.arXiv preprint arXiv:2301.04104, 2023.
- Hafner etal. (2024)Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy Lillicrap.Mastering diverse domains through world models, 2024.URL https://arxiv.org/abs/2301.04104.
- Hansen etal. (2024)Nicklas Hansen, Hao Su, and Xiaolong Wang.Td-mpc2: Scalable, robust world models for continuous control, 2024.
- He etal. (2023)Yuze He, Yushi Bai, Matthieu Lin, Wang Zhao, Yubin Hu, Jenny Sheng, Ran Yi, Juanzi Li, and Yong-Jin Liu.T3bench: Benchmarking current progress in text-to-3d generation, 2023.
- Heusel etal. (2017)Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter.Gans trained by a two time-scale update rule converge to a local nash equilibrium.Advances in neural information processing systems, 30, 2017.
- Hua etal. (2024)PuHua, Minghuan Liu, Annabella Macaluso, Yunfeng Lin, Weinan Zhang, Huazhe Xu, and Lirui Wang.Gensim2: Scaling robot data generation with multi-modal and reasoning llms, 2024.URL https://arxiv.org/abs/2410.03645.
- Huang etal. (2023a)Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu.A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions, 2023a.
- Huang etal. (2022)Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng, Jonathan Tompson, Igor Mordatch, Yevgen Chebotar, Pierre Sermanet, Noah Brown, Tomas Jackson, Linda Luu, Sergey Levine, Karol Hausman, and Brian Ichter.Inner monologue: Embodied reasoning through planning with language models, 2022.
- Huang etal. (2023b)Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and LiFei-Fei.Voxposer: Composable 3d value maps for robotic manipulation with language models.arXiv preprint arXiv:2307.05973, 2023b.
- Huang etal. (2024)Wenlong Huang, Chen Wang, Yunzhu Li, Ruohan Zhang, and LiFei-Fei.Rekep: Spatio-temporal reasoning of relational keypoint constraints for robotic manipulation.arXiv preprint arXiv:2409.01652, 2024.
- Huang etal. (2021)Zhiao Huang, Yuanming Hu, Tao Du, Siyuan Zhou, Hao Su, JoshuaB. Tenenbaum, and Chuang Gan.Plasticinelab: A soft-body manipulation benchmark with differentiable physics, 2021.
- Huang etal. (2023c)Zhiao Huang, Feng Chen, Yewen Pu, Chunru Lin, Hao Su, and Chuang Gan.Diffvl: Scaling up soft body manipulation using vision-language driven differentiable physics, 2023c.
- Katara etal. (2023)Pushkal Katara, Zhou Xian, and Katerina Fragkiadaki.Gen2sim: Scaling up robot learning in simulation with generative models, 2023.
- Li etal. (2023a)Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi.Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models, 2023a.
- Li etal. (2023b)Ming Li, Pan Zhou, Jia-Wei Liu, Jussi Keppo, Min Lin, Shuicheng Yan, and Xiangyu Xu.Instant3d: Instant text-to-3d generation, 2023b.
- Lin etal. (2021)Xingyu Lin, Yufei Wang, Jake Olkin, and David Held.Softgym: Benchmarking deep reinforcement learning for deformable object manipulation.In Jens Kober, Fabio Ramos, and Claire Tomlin (eds.), Proceedings of the 2020 Conference on Robot Learning, volume 155 of Proceedings of Machine Learning Research, pp. 432–448. PMLR, 16–18 Nov 2021.URL https://proceedings.mlr.press/v155/lin21a.html.
- Liu etal. (2024)Haotian Liu, Chunyuan Li, Yuheng Li, and YongJae Lee.Improved baselines with visual instruction tuning, 2024.URL https://arxiv.org/abs/2310.03744.
- Liu etal. (2023)Ruoshi Liu, Rundi Wu, BasileVan Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick.Zero-1-to-3: Zero-shot one image to 3d object, 2023.
- Ma etal. (2023)YechengJason Ma, William Liang, Guanzhi Wang, De-An Huang, Osbert Bastani, Dinesh Jayaraman, Yuke Zhu, Linxi Fan, and Anima Anandkumar.Eureka: Human-level reward design via coding large language models.arXiv preprint arXiv:2310.12931, 2023.
- Mohiuddin etal. (2024)Rafay Mohiuddin, SaiManoj Prakhya, Fiona Collins, Ziyuan Liu, and André Borrmann.Opensu3d: Open world 3d scene understanding using foundation models.arXiv preprint arXiv:2407.14279, 2024.
- Mu etal. (2021)Tongzhou Mu, Zhan Ling, Fanbo Xiang, Derek Yang, Xuanlin Li, Stone Tao, Zhiao Huang, Zhiwei Jia, and Hao Su.Maniskill: Generalizable manipulation skill benchmark with large-scale demonstrations, 2021.
- OpenAI (2023)OpenAI.Gpt-4 technical report, 2023.
- Pathak etal. (2017)Deepak Pathak, Pulkit Agrawal, AlexeiA Efros, and Trevor Darrell.Curiosity-driven exploration by self-supervised prediction.In International conference on machine learning, pp. 2778–2787. PMLR, 2017.
- Reimers & Gurevych (2019)Nils Reimers and Iryna Gurevych.Sentence-bert: Sentence embeddings using siamese bert-networks.In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, 11 2019.URL https://arxiv.org/abs/1908.10084.
- Reimers & Gurevych (2020)Nils Reimers and Iryna Gurevych.Making monolingual sentence embeddings multilingual using knowledge distillation.In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, 11 2020.URL https://arxiv.org/abs/2004.09813.
- Rozière etal. (2023)Baptiste Rozière, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, XiaoqingEllen Tan, Yossi Adi, Jingyu Liu, Tal Remez, Jérémy Rapin, etal.Code llama: Open foundation models for code.arXiv preprint arXiv:2308.12950, 2023.
- Salimans etal. (2016)Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and XiChen.Improved techniques for training gans.Advances in neural information processing systems, 29, 2016.
- Srivastava etal. (2021)Sanjana Srivastava, Chengshu Li, Michael Lingelbach, Roberto Martín-Martín, Fei Xia, Kent Vainio, Zheng Lian, Cem Gokmen, Shyamal Buch, C.Karen Liu, Silvio Savarese, Hyowon Gweon, Jiajun Wu, and LiFei-Fei.Behavior: Benchmark for everyday household activities in virtual, interactive, and ecological environments, 2021.
- Team etal. (2023)Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, AndrewM. Dai, Anja Hauth, Katie Millican, David Silver, Slav Petrov, Melvin Johnson, Ioannis Antonoglou, Julian Schrittwieser, Amelia Glaese, Jilin Chen, Emily Pitler, Timothy Lillicrap, Angeliki Lazaridou, Orhan Firat, James Molloy, Michael Isard, PaulR. Barham, Tom Hennigan, Benjamin Lee, Fabio Viola, Malcolm Reynolds, Yuanzhong Xu, Ryan Doherty, Eli Collins, Clemens Meyer, Eliza Rutherford, Erica Moreira, Kareem Ayoub, Megha Goel, George Tucker, Enrique Piqueras, Maxim Krikun, Iain Barr, Nikolay Savinov, Ivo Danihelka, Becca Roelofs, Anaïs White, Anders Andreassen, Tamara von Glehn, Lakshman Yagati, Mehran Kazemi, Lucas Gonzalez, Misha Khalman, Jakub Sygnowski, Alexandre Frechette, Charlotte Smith, Laura Culp, Lev Proleev, YiLuan, XiChen, James Lottes, Nathan Schucher, Federico Lebron, Alban Rrustemi, Natalie Clay, Phil Crone, Tomas Kocisky, Jeffrey Zhao, Bartek Perz, Dian Yu,Heidi Howard, Adam Bloniarz, JackW. Rae, Han Lu, Laurent Sifre, Marcello Maggioni, Fred Alcober, Dan Garrette, Megan Barnes, Shantanu Thakoor, Jacob Austin, Gabriel Barth-Maron, William Wong, Rishabh Joshi, Rahma Chaabouni, Deeni Fatiha, Arun Ahuja, Ruibo Liu, Yunxuan Li, Sarah Cogan, Jeremy Chen, Chao Jia, Chenjie Gu, Qiao Zhang, Jordan Grimstad, AleJakse Hartman, Martin Chadwick, GauravSingh Tomar, Xavier Garcia, Evan Senter, Emanuel Taropa, ThanumalayanSankaranarayana Pillai, Jacob Devlin, Michael Laskin, Diego deLasCasas, Dasha Valter, Connie Tao, Lorenzo Blanco, AdriàPuigdomènech Badia, David Reitter, Mianna Chen, Jenny Brennan, Clara Rivera, Sergey Brin, Shariq Iqbal, Gabriela Surita, Jane Labanowski, Abhi Rao, Stephanie Winkler, Emilio Parisotto, Yiming Gu, Kate Olszewska, Yujing Zhang, Ravi Addanki, Antoine Miech, Annie Louis, LaurentEl Shafey, Denis Teplyashin, Geoff Brown, Elliot Catt, Nithya Attaluri, Jan Balaguer, Jackie Xiang, Pidong Wang, Zoe Ashwood, Anton Briukhov, Albert Webson,Sanjay Ganapathy, Smit Sanghavi, Ajay Kannan, Ming-Wei Chang, Axel Stjerngren, Josip Djolonga, Yuting Sun, Ankur Bapna, Matthew Aitchison, Pedram Pejman, Henryk Michalewski, Tianhe Yu, Cindy Wang, Juliette Love, Junwhan Ahn, Dawn Bloxwich, Kehang Han, Peter Humphreys, Thibault Sellam, James Bradbury, Varun Godbole, Sina Samangooei, Bogdan Damoc, Alex Kaskasoli, Sébastien M.R. Arnold, Vijay Vasudevan, Shubham Agrawal, Jason Riesa, Dmitry Lepikhin, Richard Tanburn, Srivatsan Srinivasan, Hyeontaek Lim, Sarah Hodkinson, Pranav Shyam, Johan Ferret, Steven Hand, Ankush Garg, TomLe Paine, Jian Li, Yujia Li, Minh Giang, Alexander Neitz, Zaheer Abbas, Sarah York, Machel Reid, Elizabeth Cole, Aakanksha Chowdhery, Dipanjan Das, Dominika Rogozińska, Vitaly Nikolaev, Pablo Sprechmann, Zachary Nado, Lukas Zilka, Flavien Prost, Luheng He, Marianne Monteiro, Gaurav Mishra, Chris Welty, Josh Newlan, Dawei Jia, Miltiadis Allamanis, ClaraHuiyi Hu, Raoul deLiedekerke, Justin Gilmer, Carl Saroufim, Shruti Rijhwani, ShaoboHou, Disha Shrivastava, Anirudh Baddepudi, Alex Goldin, Adnan Ozturel, Albin Cassirer, Yunhan Xu, Daniel Sohn, Devendra Sachan, ReinaldKim Amplayo, Craig Swanson, Dessie Petrova, Shashi Narayan, Arthur Guez, Siddhartha Brahma, Jessica Landon, Miteyan Patel, Ruizhe Zhao, Kevin Villela, Luyu Wang, Wenhao Jia, Matthew Rahtz, Mai Giménez, Legg Yeung, Hanzhao Lin, James Keeling, Petko Georgiev, Diana Mincu, Boxi Wu, Salem Haykal, Rachel Saputro, Kiran Vodrahalli, James Qin, Zeynep Cankara, Abhanshu Sharma, Nick Fernando, Will Hawkins, Behnam Neyshabur, Solomon Kim, Adrian Hutter, Priyanka Agrawal, Alex Castro-Ros, George vanden Driessche, Tao Wang, Fan Yang, Shuo yiin Chang, Paul Komarek, Ross McIlroy, Mario Lučić, Guodong Zhang, Wael Farhan, Michael Sharman, Paul Natsev, Paul Michel, Yong Cheng, Yamini Bansal, Siyuan Qiao, Kris Cao, Siamak Shakeri, Christina Butterfield, Justin Chung, PaulKishan Rubenstein, Shivani Agrawal, Arthur Mensch, Kedar Soparkar, Karel Lenc, Timothy Chung, Aedan Pope, LorenMaggiore, Jackie Kay, Priya Jhakra, Shibo Wang, Joshua Maynez, Mary Phuong, Taylor Tobin, Andrea Tacchetti, Maja Trebacz, Kevin Robinson, Yash Katariya, Sebastian Riedel, Paige Bailey, Kefan Xiao, Nimesh Ghelani, Lora Aroyo, Ambrose Slone, Neil Houlsby, Xuehan Xiong, Zhen Yang, Elena Gribovskaya, Jonas Adler, Mateo Wirth, Lisa Lee, Music Li, Thais Kagohara, Jay Pavagadhi, Sophie Bridgers, Anna Bortsova, Sanjay Ghemawat, Zafarali Ahmed, Tianqi Liu, Richard Powell, Vijay Bolina, Mariko Iinuma, Polina Zablotskaia, James Besley, Da-Woon Chung, Timothy Dozat, Ramona Comanescu, Xiance Si, Jeremy Greer, Guolong Su, Martin Polacek, RaphaëlLopez Kaufman, Simon Tokumine, Hexiang Hu, Elena Buchatskaya, Yingjie Miao, Mohamed Elhawaty, Aditya Siddhant, Nenad Tomasev, Jinwei Xing, Christina Greer, Helen Miller, Shereen Ashraf, Aurko Roy, Zizhao Zhang, Ada Ma, Angelos Filos, Milos Besta, Rory Blevins, Ted Klimenko, Chih-Kuan Yeh, Soravit Changpinyo, Jiaqi Mu, Oscar Chang, Mantas Pajarskas, Carrie Muir, Vered Cohen,CharlineLe Lan, Krishna Haridasan, Amit Marathe, Steven Hansen, Sholto Douglas, Rajkumar Samuel, Mingqiu Wang, Sophia Austin, Chang Lan, Jiepu Jiang, Justin Chiu, JaimeAlonso Lorenzo, LarsLowe Sjösund, Sébastien Cevey, Zach Gleicher, Thi Avrahami, Anudhyan Boral, Hansa Srinivasan, Vittorio Selo, Rhys May, Konstantinos Aisopos, Léonard Hussenot, LivioBaldini Soares, Kate Baumli, MichaelB. Chang, Adrià Recasens, Ben Caine, Alexander Pritzel, Filip Pavetic, Fabio Pardo, Anita Gergely, Justin Frye, Vinay Ramasesh, Dan Horgan, Kartikeya Badola, Nora Kassner, Subhrajit Roy, Ethan Dyer, Víctor Campos, Alex Tomala, Yunhao Tang, DaliaEl Badawy, Elspeth White, Basil Mustafa, Oran Lang, Abhishek Jindal, Sharad Vikram, Zhitao Gong, Sergi Caelles, Ross Hemsley, Gregory Thornton, Fangxiaoyu Feng, Wojciech Stokowiec, CeZheng, Phoebe Thacker, Çağlar Ünlü, Zhishuai Zhang, Mohammad Saleh, James Svensson, Max Bileschi, Piyush Patil, Ankesh Anand, Roman Ring, Katerina Tsihlas, Arpi Vezer, Marco Selvi, TobyShevlane, Mikel Rodriguez, Tom Kwiatkowski, Samira Daruki, Keran Rong, Allan Dafoe, Nicholas FitzGerald, Keren Gu-Lemberg, Mina Khan, LisaAnne Hendricks, Marie Pellat, Vladimir Feinberg, James Cobon-Kerr, Tara Sainath, Maribeth Rauh, SayedHadi Hashemi, Richard Ives, Yana Hasson, YaGuang Li, Eric Noland, Yuan Cao, Nathan Byrd, LeHou, Qingze Wang, Thibault Sottiaux, Michela Paganini, Jean-Baptiste Lespiau, Alexandre Moufarek, Samer Hassan, Kaushik Shivakumar, Joost van Amersfoort, Amol Mandhane, Pratik Joshi, Anirudh Goyal, Matthew Tung, Andrew Brock, Hannah Sheahan, Vedant Misra, Cheng Li, Nemanja Rakićević, Mostafa Dehghani, Fangyu Liu, Sid Mittal, Junhyuk Oh, Seb Noury, Eren Sezener, Fantine Huot, Matthew Lamm, NicolaDe Cao, Charlie Chen, Gamaleldin Elsayed, EdChi, Mahdis Mahdieh, Ian Tenney, Nan Hua, Ivan Petrychenko, Patrick Kane, Dylan Scandinaro, Rishub Jain, Jonathan Uesato, Romina Datta, Adam Sadovsky, Oskar Bunyan, Dominik Rabiej, Shimu Wu, John Zhang, Gautam Vasudevan, Edouard Leurent,Mahmoud Alnahlawi, Ionut Georgescu, Nan Wei, Ivy Zheng, Betty Chan, PamG Rabinovitch, Piotr Stanczyk, YeZhang, David Steiner, Subhajit Naskar, Michael Azzam, Matthew Johnson, Adam Paszke, Chung-Cheng Chiu, JaumeSanchez Elias, Afroz Mohiuddin, Faizan Muhammad, Jin Miao, Andrew Lee, Nino Vieillard, Sahitya Potluri, Jane Park, Elnaz Davoodi, Jiageng Zhang, Jeff Stanway, Drew Garmon, Abhijit Karmarkar, Zhe Dong, Jong Lee, Aviral Kumar, Luowei Zhou, Jonathan Evens, William Isaac, Zhe Chen, Johnson Jia, Anselm Levskaya, Zhenkai Zhu, Chris Gorgolewski, Peter Grabowski, YuMao, Alberto Magni, Kaisheng Yao, Javier Snaider, Norman Casagrande, Paul Suganthan, Evan Palmer, Geoffrey Irving, Edward Loper, Manaal Faruqui, Isha Arkatkar, Nanxin Chen, Izhak Shafran, Michael Fink, Alfonso Castaño, Irene Giannoumis, Wooyeol Kim, Mikołaj Rybiński, Ashwin Sreevatsa, Jennifer Prendki, David Soergel, Adrian Goedeckemeyer, Willi Gierke, Mohsen Jafari, Meenu Gaba, Jeremy Wiesner, DianaGage Wright, Yawen Wei, Harsha Vashisht,Yana Kulizhskaya, Jay Hoover, Maigo Le, LuLi, Chimezie Iwuanyanwu, LuLiu, Kevin Ramirez, Andrey Khorlin, Albert Cui, Tian LIN, Marin Georgiev, Marcus Wu, Ricardo Aguilar, Keith Pallo, Abhishek Chakladar, Alena Repina, Xihui Wu, Tom vander Weide, Priya Ponnapalli, Caroline Kaplan, Jiri Simsa, Shuangfeng Li, Olivier Dousse, Fan Yang, Jeff Piper, Nathan Ie, Minnie Lui, Rama Pasumarthi, Nathan Lintz, Anitha Vijayakumar, LamNguyen Thiet, Daniel Andor, Pedro Valenzuela, Cosmin Paduraru, Daiyi Peng, Katherine Lee, Shuyuan Zhang, Somer Greene, DucDung Nguyen, Paula Kurylowicz, Sarmishta Velury, Sebastian Krause, Cassidy Hardin, Lucas Dixon, Lili Janzer, Kiam Choo, Ziqiang Feng, Biao Zhang, Achintya Singhal, Tejasi Latkar, Mingyang Zhang, Quoc Le, ElenaAllica Abellan, Dayou Du, Dan McKinnon, Natasha Antropova, Tolga Bolukbasi, Orgad Keller, David Reid, Daniel Finchelstein, MariaAbi Raad, Remi Crocker, Peter Hawkins, Robert Dadashi, Colin Gaffney, Sid Lall, Ken Franko, Egor Filonov, Anna Bulanova, RémiLeblond, Vikas Yadav, Shirley Chung, Harry Askham, LuisC. Cobo, Kelvin Xu, Felix Fischer, Jun Xu, Christina Sorokin, Chris Alberti, Chu-Cheng Lin, Colin Evans, Hao Zhou, Alek Dimitriev, Hannah Forbes, Dylan Banarse, Zora Tung, Jeremiah Liu, Mark Omernick, Colton Bishop, Chintu Kumar, Rachel Sterneck, Ryan Foley, Rohan Jain, Swaroop Mishra, Jiawei Xia, Taylor Bos, Geoffrey Cideron, Ehsan Amid, Francesco Piccinno, Xingyu Wang, Praseem Banzal, Petru Gurita, Hila Noga, Premal Shah, DanielJ. Mankowitz, Alex Polozov, Nate Kushman, Victoria Krakovna, Sasha Brown, MohammadHossein Bateni, Dennis Duan, Vlad Firoiu, Meghana Thotakuri, Tom Natan, Anhad Mohananey, Matthieu Geist, Sidharth Mudgal, Sertan Girgin, Hui Li, Jiayu Ye, Ofir Roval, Reiko Tojo, Michael Kwong, James Lee-Thorp, Christopher Yew, Quan Yuan, Sumit Bagri, Danila Sinopalnikov, Sabela Ramos, John Mellor, Abhishek Sharma, Aliaksei Severyn, Jonathan Lai, Kathy Wu, Heng-Tze Cheng, David Miller, Nicolas Sonnerat, Denis Vnukov, Rory Greig, JenniferBeattie, Emily Caveness, Libin Bai, Julian Eisenschlos, Alex Korchemniy, Tomy Tsai, Mimi Jasarevic, Weize Kong, Phuong Dao, Zeyu Zheng, Frederick Liu, Fan Yang, Rui Zhu, Mark Geller, TianHuey Teh, Jason Sanmiya, Evgeny Gladchenko, Nejc Trdin, Andrei Sozanschi, Daniel Toyama, Evan Rosen, Sasan Tavakkol, Linting Xue, Chen Elkind, Oliver Woodman, John Carpenter, George Papamakarios, Rupert Kemp, Sushant Kafle, Tanya Grunina, Rishika Sinha, Alice Talbert, Abhimanyu Goyal, Diane Wu, Denese Owusu-Afriyie, Cosmo Du, Chloe Thornton, Jordi Pont-Tuset, Pradyumna Narayana, Jing Li, Sabaer Fatehi, John Wieting, Omar Ajmeri, Benigno Uria, Tao Zhu, Yeongil Ko, Laura Knight, Amélie Héliou, Ning Niu, Shane Gu, Chenxi Pang, Dustin Tran, Yeqing Li, Nir Levine, Ariel Stolovich, Norbert Kalb, Rebeca Santamaria-Fernandez, Sonam Goenka, Wenny Yustalim, Robin Strudel, Ali Elqursh, Balaji Lakshminarayanan, Charlie Deck, Shyam Upadhyay, Hyo Lee, Mike Dusenberry, Zonglin Li, Xuezhi Wang, Kyle Levin, Raphael Hoffmann, DanHoltmann-Rice, Olivier Bachem, Summer Yue, Sho Arora, Eric Malmi, Daniil Mirylenka, Qijun Tan, Christy Koh, SoheilHassas Yeganeh, Siim Põder, Steven Zheng, Francesco Pongetti, Mukarram Tariq, Yanhua Sun, Lucian Ionita, Mojtaba Seyedhosseini, Pouya Tafti, Ragha Kotikalapudi, Zhiyu Liu, Anmol Gulati, Jasmine Liu, Xinyu Ye, Bart Chrzaszcz, Lily Wang, Nikhil Sethi, Tianrun Li, Ben Brown, Shreya Singh, Wei Fan, Aaron Parisi, Joe Stanton, Chenkai Kuang, Vinod Koverkathu, ChristopherA. Choquette-Choo, Yunjie Li, TJLu, Abe Ittycheriah, Prakash Shroff, Pei Sun, Mani Varadarajan, Sanaz Bahargam, Rob Willoughby, David Gaddy, Ishita Dasgupta, Guillaume Desjardins, Marco Cornero, Brona Robenek, Bhavishya Mittal, Ben Albrecht, Ashish Shenoy, Fedor Moiseev, Henrik Jacobsson, Alireza Ghaffarkhah, Morgane Rivière, Alanna Walton, Clément Crepy, Alicia Parrish, Yuan Liu, Zongwei Zhou, Clement Farabet, Carey Radebaugh, Praveen Srinivasan, Claudia vander Salm, Andreas Fidjeland, Salvatore Scellato, Eri Latorre-Chimoto,Hanna Klimczak-Plucińska, David Bridson, Dario deCesare, Tom Hudson, Piermaria Mendolicchio, Lexi Walker, Alex Morris, Ivo Penchev, Matthew Mauger, Alexey Guseynov, Alison Reid, Seth Odoom, Lucia Loher, Victor Cotruta, Madhavi Yenugula, Dominik Grewe, Anastasia Petrushkina, Tom Duerig, Antonio Sanchez, Steve Yadlowsky, Amy Shen, Amir Globerson, Adam Kurzrok, Lynette Webb, Sahil Dua, Dong Li, Preethi Lahoti, Surya Bhupatiraju, Dan Hurt, Haroon Qureshi, Ananth Agarwal, Tomer Shani, Matan Eyal, Anuj Khare, ShreyasRammohan Belle, Lei Wang, Chetan Tekur, MihirSanjay Kale, Jinliang Wei, Ruoxin Sang, Brennan Saeta, Tyler Liechty, YiSun, Yao Zhao, Stephan Lee, Pandu Nayak, Doug Fritz, ManishReddy Vuyyuru, John Aslanides, Nidhi Vyas, Martin Wicke, Xiao Ma, Taylan Bilal, Evgenii Eltyshev, Daniel Balle, Nina Martin, Hardie Cate, James Manyika, Keyvan Amiri, Yelin Kim, XiXiong, Kai Kang, Florian Luisier, Nilesh Tripuraneni, David Madras, Mandy Guo, Austin Waters, Oliver Wang, Joshua Ainslie, Jason Baldridge, HanZhang, Garima Pruthi, Jakob Bauer, Feng Yang, Riham Mansour, Jason Gelman, Yang Xu, George Polovets, JiLiu, Honglong Cai, Warren Chen, XiangHai Sheng, Emily Xue, Sherjil Ozair, Adams Yu, Christof Angermueller, Xiaowei Li, Weiren Wang, Julia Wiesinger, Emmanouil Koukoumidis, Yuan Tian, Anand Iyer, Madhu Gurumurthy, Mark Goldenson, Parashar Shah, MKBlake, Hongkun Yu, Anthony Urbanowicz, Jennimaria Palomaki, Chrisantha Fernando, Kevin Brooks, Ken Durden, Harsh Mehta, Nikola Momchev, Elahe Rahimtoroghi, Maria Georgaki, Amit Raul, Sebastian Ruder, Morgan Redshaw, Jinhyuk Lee, Komal Jalan, Dinghua Li, Ginger Perng, Blake Hechtman, Parker Schuh, Milad Nasr, Mia Chen, Kieran Milan, Vladimir Mikulik, Trevor Strohman, Juliana Franco, Tim Green, Demis Hassabis, Koray Kavukcuoglu, Jeffrey Dean, and Oriol Vinyals.Gemini: A family of highly capable multimodal models, 2023.
- Tobin etal. (2017)Joshua Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and P.Abbeel.Domain randomization for transferring deep neural networks from simulation to the real world.2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 23–30, 2017.URL https://api.semanticscholar.org/CorpusID:2413610.
- Tong etal. (2024a)Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, SaiCharitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, Austin Wang, Rob Fergus, Yann LeCun, and Saining Xie.Cambrian-1: A fully open, vision-centric exploration of multimodal llms, 2024a.URL https://arxiv.org/abs/2406.16860.
- Tong etal. (2024b)Shengbang Tong, Zhuang Liu, Yuexiang Zhai, YiMa, Yann LeCun, and Saining Xie.Eyes wide shut? exploring the visual shortcomings of multimodal llms, 2024b.URL https://arxiv.org/abs/2401.06209.
- Touvron etal. (2023)Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, etal.Llama: Open and efficient foundation language models.arXiv preprint arXiv:2302.13971, 2023.
- Wang etal. (2018)Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and SamuelR Bowman.Glue: A multi-task benchmark and analysis platform for natural language understanding.arXiv preprint arXiv:1804.07461, 2018.
- Wang etal. (2024)Lirui Wang, Yiyang Ling, Zhecheng Yuan, Mohit Shridhar, Chen Bao, Yuzhe Qin, Bailin Wang, Huazhe Xu, and Xiaolong Wang.Gensim: Generating robotic simulation tasks via large language models, 2024.
- Wang etal. (2023a)Yufei Wang, Zhanyi Sun, Zackory Erickson, and David Held.One policy to dress them all: Learning to dress people with diverse poses and garments, 2023a.
- Wang etal. (2023b)Yufei Wang, Zhou Xian, Feng Chen, Tsun-Hsuan Wang, Yian Wang, Zackory Erickson, David Held, and Chuang Gan.Robogen: Towards unleashing infinite data for automated robot learning via generative simulation, 2023b.
- Xu etal. (2023)Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong.Imagereward: Learning and evaluating human preferences for text-to-image generation, 2023.
- Yang etal. (2024)Sizhe Yang, Qian Luo, Anumpam Pani, and Yanchao Yang.Bbsea: An exploration of brain-body synchronization for embodied agents.arXiv preprint arXiv:2402.08212, 2024.
- Yu etal. (2020)Tianhe Yu, Deirdre Quillen, Zhanpeng He, Ryan Julian, Karol Hausman, Chelsea Finn, and Sergey Levine.Meta-world: A benchmark and evaluation for multi-task and meta reinforcement learning.In Conference on robot learning, pp. 1094–1100. PMLR, 2020.
- Yuan etal. (2024)Zhecheng Yuan, Sizhe Yang, PuHua, Can Chang, Kaizhe Hu, and Huazhe Xu.Rl-vigen: A reinforcement learning benchmark for visual generalization.Advances in Neural Information Processing Systems, 36, 2024.
- Zhang etal. (2023a)Xinlu Zhang, Yujie Lu, Weizhi Wang, AnYan, Jun Yan, Lianke Qin, Heng Wang, Xifeng Yan, WilliamYang Wang, and LindaRuth Petzold.Gpt-4v(ision) as a generalist evaluator for vision-language tasks, 2023a.
- Zhang etal. (2023b)Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, YuZhang, Yulong Chen, Longyue Wang, AnhTuan Luu, Wei Bi, Freda Shi, and Shuming Shi.Siren’s song in the ai ocean: A survey on hallucination in large language models, 2023b.
- Zhu etal. (2018)Yaoming Zhu, Sidi Lu, Lei Zheng, Jiaxian Guo, Weinan Zhang, Jun Wang, and Yong Yu.Texygen: A benchmarking platform for text generation models.The 41st International ACM SIGIR Conference on Research & Development in Information Retrieval, 2018.URL https://api.semanticscholar.org/CorpusID:3636178.
Appendix A Appendix
A.1 Prompt for evaluation
Here is the prompt for the evaluation of the completion score:
⬇
You are an assessment expert responsible for the Task Completion rate. Your task is to score the completion rate for the task in the following rules:
1. Evaluate the completion rate for the robotics task.
2. During the evaluation, you will receive 8 images and a basic description of the task.
3. During the evaluation, you need to make very careful judgments and evaluate the completion of the task based on the order of the pictures and the task description.
4. In the evaluation you need to pay attention to the smoothness of the trajectory.
5. Assign a score between 0 and 10, with 10 being the highest. Do not provide a complete answer.
6. Your should provide the answer in the following format:
Score: X
Here is the prompt for the evaluation of the scene alignment score without the caption model:
⬇
You are an assessment expert responsible for Task description and Scene images pairs. Your task is to score the Scene caption according to the following requirements:
1. Evaluate how well the Scene images covers the scene of the robotics task. You should consider whether the scene is similar to the requirement of the task.
2. During the evaluation, you will receive 4 images and a basic description of the task.
3. In the evaluation, you should pay attention to the alignment between the Scene image and the real-world task following the description of the task.
4. A good scene should not only provide an environment for completing a robotics task but should also contain items that may appear near the task, even though they may have nothing to do with the task itself.
5. Assign a score between 1 and 5, with 5 being the highest.
6. Your should provide the answer in the following format:
Score: X
Here is the prompt for the evaluation of the scene alignment score with the caption model:
⬇
You are an assessment expert responsible for Task description and Scene captions pairs. Your task is to score the Scene caption according to the following requirements:
1. Evaluate how well the Scene captions covers the scene of the robotics task. You should consider whether the scene is similar to the requirement of the task.
2. In the evaluation, you should pay attention to the alignment between the Scene captions and the real-world task following the description of the task.
3. A good scene should not only provide an environment for completing a robotics task but should also contain items that may appear near the task, even though they may have nothing to do with the task itself.
4. Scene caption sets will be a set of different views to the scene.
5. Assign a score between 1 and 5, with 5 being the highest. Do not provide a complete answer; give the score in the format: Score: 3
A.2 Human Evaluation Result
In table LABEL:tab:alignment and table LABEL:tab:completion, we collected the scoring results from 18 researchers in the field of robotics on 20 tasks, which were sourced from RoboGen and GenSim, respectively. We conducted five evaluations for each model, with the values in parentheses representing the variance of the five evaluations. Based on these two tables, we derived the results presented in Figure 3.
Task Name (RoboGen)HumanGPT-4+Blip2CambrianLLavaGPT-4vGPT-4+LLavaGPT-4+CambrianOpen Laptop3.62.56(0.9)2.00(0)3.00(0)3.96(0.03)3.20(1.20)2.00(0.00)Change Lamp Direction3.452.88(0.75)2.00(0)3.00(0)3.96(0.03)3.80(0.08)3.28(0.21)Flush the Toilet3.553.00(0.91)0.00(0)3.00(0)3.40(0.02)2.80(1.20)4.00(0)Extend Display Screen2.552.16(0.13)0.00(0)4.00(0)3.80(0.14)2.00(0)1.96(0.01)Close the Oven Door3.451.80(0.2)0.00(0)3.00(0)3.60(0.04)2.00(0)2.00(0)Set Oven Timer3.202.80(1.2)0.00(0)3.00(0)3.04(0.03)3.96(0.01)4.00(0)Close Window3.102.40(0.48)1.00(0)3.00(0)3.96(0.01)1.60(0.30)1.00(0)Adjust Water Flow2.851.40(0.3)2.00(0)3.00(0)3.36(0.03)1.24(0.11)1.48(0.05)Open Both Table Doors3.202.24(0.29)5.00(0)3.00(0)3.24(0.05)2.00(0)2.00(0)Press Start Button3.203.24(1.29)0.00(0)3.00(0)3.48(0.03)2.20(2.70)4.00(0)Task Name (Gensim)HumanGPT-4+Blip2CambrianLLavaGPT-4vGPT-4+LLavaGPT-4+CambrianAlign Balls in Colored Boxes4.304.00(0.00)2.00(0)3.00(0)3.36(0.07)1.33(0.33)2.00(0)Block Pyramid with Limited Space4.104.00(3.00)2.00(0)3.00(0)3.64(0.07)1.50(0.50)2.04(0.01)Align Spheres in Colored Zones4.204.40(0.16)2.00(0)3.40(0.30)4.32(0.03)2.00(0.00)2.00(0)Color Coded Blocks on Corner3.154.13(0.05)0.40(0.80)3.00(0)3.68(0.07)2.00(0)3.24(0.61)Align Rope Cross Zone4.353.53(0.65)2.00(0)1.00(0)3.56(0.11)1.00(0)1.80(0.20)Color Ordered Insertion4.402.00(0.00)2.00(0)3.00(0)4.08(0.11)2.00(0)2.00(0)Color Specific Container Fill4.252.00(0.00)0.60(0.30)3.00(0)3.84(0.03)2.00(0)2.20(0.20)Color Coordinated Zone Stacking3.704.27(0.21)2.00(0)3.00(0)3.88(0.03)1.00(0)2.40(0.80)Vertical Insertion Blocks3.753.13(1.77)1.60(0.80)3.00(0)3.44(0.09)2.07(1.21)2.20(0.20)Color Blocks in Cylinder Maze2.652.47(0.65)0.00(0)3.00(0)3.12(0.05)2.00(0)1.76(0.13)
Task Name (RoboGen) | Human | Cambrian-8B | LLava-1.5 | GPT-4 |
---|---|---|---|---|
Open Laptop | 8.2 | 0.00(0) | 8.00(0) | 7.96(0.01) |
Change Lamp Direction | 8.4 | 0.00(0) | 8.00(0) | 6.40(0.92) |
Flush the Toilet | 8.3 | 0.00(0) | 8.00(0) | 7.40(0.06) |
Extend Display Screen | 5.1 | 0.00(0) | 8.00(0) | 6.36(1.11) |
Close the Oven Door | 4.8 | 0.00(0) | 8.00(0) | 5.80(1.02) |
Set Oven Timer | 7.8 | 0.00(0) | 8.00(0) | 6.40(0.04) |
Close Window | 9.3 | 0.00(0) | 8.00(0) | 6.80(0.42) |
Adjust Water Flow | 7.2 | 0.00(0) | 8.00(0) | 7.64(0.07) |
Open Both Table Doors | 6.3 | 0.00(0) | 8.00(0) | 5.76(1.35) |
Press Start Button | 8.4 | 0.00(0) | 8.00(0) | 6.88(0.81) |
Task Name (Gensim) | Human | Cambrian-8B | LLava-1.5 | GPT-4 |
Align Balls in Colored Boxes | 9.2 | 0.00(0) | 8.00(0) | 4.92(0.23) |
Block Pyramid with Limited Space | 6.8 | 0.00(0) | 8.00(0) | 5.16(2.35) |
Align Spheres in Colored Zones | 4.9 | 0.00(0) | 8.00(0) | 3.16(2.97) |
Color Coded Blocks on Corner | 6.2 | 0.00(0) | 8.00(0) | 6.48(1.27) |
Align Rope Cross Zone | 8.7 | 0.80(3.20) | 8.00(0) | 4.25(2.81) |
Color Ordered Insertion | 9.5 | 1.60(4.80) | 8.00(0) | 5.56(1.21) |
Color Specific Container Fill | 9.2 | 0.80(3.20) | 8.00(0) | 5.88(0.43) |
Color Coordinated Zone Stacking | 8.4 | 0.00(0) | 8.00(0) | 5.52(0.83) |
Vertical Insertion Blocks | 8.1 | 0.00(0) | 8.00(0) | 3.72(1.31) |
Color Blocks in Cylinder Maze | 6.3 | 0.00(0) | 8.00(0) | 6.08(1.69) |
A.3 Diversity and Generalization Experiment Details
A.3.1 Task Selection and Grouping
For GenSim, we use all the templates and generated tasks released by the authors. For RoboGen, we only use the manipulation tasks but not locomotion and soft body because the locomotion tasks yield very poor learning performance and the soft-body tasks are not publicly available at the time. For BBSEA, we perform generation following the instructions provided by the authors.
We group these tasks mainly for two reasons: (1) grouped tasks offer more perspectives for analysis, and (2) the latent dynamics model and diffusion policy training, with their original implementation, are insufficient for learning a large number of tasks. The dimensions to consider include scene configuration (e.g., table-top vs. ground), skill, and objects involved (e.g., pick-and-place using a suction gripper vs. moving piles of grains using a shovel-like end-effector).
A summary with examples is shown in Table6.
Project | Group | # Tasks | Desc. Examples | ||
---|---|---|---|---|---|
GenSim | Placement | 19 |
| ||
Stacking | 30 |
| |||
Piles | 11 | “sweeping-piles: push piles of small objects into a target goal zone” | |||
Assembling | 14 |
| |||
RoboGen | Table-Top | 17 |
| ||
Ground | 15 |
| |||
BBSEA | Table | 49 | “Gather all objects and organize them in the green bin” | ||
Drawer | 26 | “Open the drawer using the drawer handle” |
A.3.2 Trajectory Collection
Trajectories are collected for both dynamics model learning and imitation learning.GenSim implements an oracle agent for generating demonstrations. The oracle agent’s action specifies the target end-effector pose command, which is executed by a low-level Inverse Kinematics controller with joint commands. Therefore, we collect transitions of both high-level (target end-effector pose as actions) and low-level (joint commands as actions) with the observations being RBG image and robot joint states. We collect for each task 40 trajectories with manually varied random seeds.
RoboGen decomposes a long-horizon task into multiple stages, each solved by motion planning or reinforcement learning. In this work, we are only concerned with the sub-tasks that require reinforcement learning. Specifically, we run their pipeline to train a policy for each task and subsequently collect trajectories using that policy. The action space includes joint and gripper commands and the observation space includes RGB images and robot joint states. We collect for each task 40 trajectories with manually varied random seeds.
BBSEA generates trajectory demonstration by querying ChatGPT to output parameterized action primitives, which are then executed by low-level controllers. Since BBSEA does not have officially released tasks, we run the pipeline for generation and collection for each of the 32 scenes, giving 256 trajectories in total. BBSEA’s proposed pipeline additionally filters success trajectories. But here we use all trajectories for learning the dynamics model.
A.3.3 Dynamics Model Training Details
We adopt a popular community implementation (Hafner etal., 2024). For all experiments, the model is trained for 10 epochs with a batch size of 8. The data was chunked into sequences of size 40/40/20 for GenSim/RoboGen/BBSEA. All other hyperparameters are kept as default. Since DreamerV3 is designed for reinforcement learning from visual observations, its model architecture is expressive and robust to different domains. Data augmentation could be used to improve its robustness to aspects such as the variation in color and appearance further to obtain a lower prediction error. However, we do not incorporate that in this paper for simplicity.
A.3.4 Imitation Learning Model Training Details
For GenSim and RoboGen, the implementation is adapted from the official release of Diffusion Policy (Chi etal., 2023). We use the configuration provided by the authors of Diffusion Policy for image-state observation. For all experiments, the policy is trained for 8000 epochs. For BBSEA, we use the image-language diffusion policy implementation provided by the authors.