RoboCLIP: One Demonstration is Enough to Learn Robot Policies

Reward specification is a notoriously difficult problem in reinforcement learning, requiring extensive expert supervision to design robust reward functions. Imitation learning (IL) methods attempt to circumvent these problems by utilizing expert demonstrations but typically require a large number of in-domain expert demonstrations. Inspired by advances in the field of Video-and-Language Models (VLMs), we present RoboCLIP, an online imitation learning method that uses a single demonstration (overcoming the large data requirement) in the form of a video demonstration or a textual description of the task to generate rewards without manual reward function design. Additionally, RoboCLIP can also utilize out-of-domain demonstrations, like videos of humans solving the task for reward generation, circumventing the need to have the same demonstration and deployment domains. RoboCLIP utilizes pretrained VLMs without any finetuning for reward generation. Reinforcement learning agents trained with RoboCLIP rewards demonstrate 2-3 times higher zero-shot performance than competing imitation learning methods on downstream robot manipulation tasks, doing so using only one video/text demonstration.

Paper

References (46)

Scroll for more · 34 remaining

Similar papers

Peer review

Reviewer EmDU5/10 · confidence 4/52023-07-02

Summary

This study seeks to address the problem of data efficiency in imitation learning, aiming to train a robot to perform manipulation tasks after only one demonstration. In order to accomplish this, the research proposes using a pre-trained visual language model (VLM) to encode the demonstration video or task description text from the expert. The use of the dot product between the learnt and demonstrated embeddings would then serve as a reward function. Experiments are conducted in Metaworld and the Franka simulation, with tasks including "closing a door," "closing a drawer," and "pressing a button."

Strengths

- The idea of utilizing a pretrained visual language model to encode expert demonstrations or text descriptions as a reward function is neat and straightforward. - This work conducts an extensive amount of robot experiments in simulation including"closing a door", "pressing a button", "Slide Cabinet " and so on.

Weaknesses

- The use of dot product between expert embeddings and learned policy embeddings as the objective reward function is not adequately explained. The rationale behind this decision and its effectiveness would benefit from further elaboration. - The Domain Alignment Confusion Matrix reveals a poor similarity alignment between the text description and the video from the robot camera. This implies a potential issue, where goal descriptions in text form may not correspond accurately to the video data. - Moreover, the experimental performance demonstrated in the accompanying video is weak, with numerous instances of failure. - The selected baseline is outdated. GAIL was introduced in 2016 and AIRL was in 2017. - The proposed method fails to improve and instead performs significantly worse when presented with an increased number of expert demonstrations.

Questions

Q1: In Figure 4, what is the reason behind GAIL, AIRL, and a method called "Task Reward" reflecting an identical reward across all three tasks?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

See Weaknesses and Questions.

Reviewer S3ve6/10 · confidence 3/52023-07-06

Summary

RoboCLIP is an innovative online imitation learning method that addresses the challenging problem of reward specification in reinforcement learning. Unlike traditional approaches that rely on expert supervision and extensive data, RoboCLIP utilizes a single demonstration, either in the form of a video or a textual description, to generate rewards without the need for manual reward function design. It can even leverage out-of-domain demonstrations, eliminating the requirement for demonstrations from the same domain as deployment. Basically the method compares the embedding between the video generated by the policy and video given to it for imitation to get a reward for training an RL agent. The idea is simple and intuitive.

Strengths

Although the idea is simple, the authors verify through experimentation that learning policies from videos is possible. The results are great and contribute to advance in robot policy learning from videos.

Weaknesses

Out-of-distribution results are underwhelming. Maybe a future direction could to focus more on the object state to train better policies. No real robot results :( The whole point of something like this should be easy applicability to the real world.

Questions

I don't have any questions for the authors.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

limitations are discussed adequately.

Reviewer Dcaj6/10 · confidence 4/52023-07-09

Summary

This paper presents a novel approach to leverage pretrained Vision-Language Models (VLMs) for training Reinforcement Learning (RL) agents by utilizing them to generate rewards. Specifically, the authors employ S3D, pretrained on HowTo100M, as the backbone and measure the dissimilarity between online agent experience and task demonstrations to serve as the reward signal for training task-specific RL policies. They demonstrate the effectiveness of their proposed method on the Metaworld Environment and the Franka Kitchen Environment.

Strengths

1. This paper addresses an important and widely recognized problem, which is the grounding of VLMs. 2. The idea behind the proposed method is straightforward and intuitive. 3. The author thoroughly examines the performance of the proposed approach from six different perspectives: alignment, text-to-reward mapping, video-to-reward mapping, generalization to out-of-domain videos, decompositional generalizability, and an ablation study.

Weaknesses

I have some major concerns about the experiment section of this paper. There are several results that are not convincing and can be quite confusing. 1. The domain alignment between the demonstration and text description, as mentioned by the author, is indeed unsatisfactory. For instance, the policy should learn to close the green drawer instead of opening it when the task descriptor is "open green drawer." Using this misleading reward can lead to misleading learning outcomes. 2. In the language-to-reward section (Section 4.2), the author only demonstrates the performance on three "easy" tasks, which is not sufficiently convincing. Additionally, it is unclear why the proposed method outperforms dense task rewards. The dense task reward can be considered an oracle, so it raises questions about whether the performance improvement is due to insufficient training epochs or simply because the rewards themselves are inadequate. 3. The qualitative results presented in Section 4.3 are challenging to comprehend. It is difficult to understand how the robot can open the cabinet without even touching it. While I can imagine what happened in between, the visualizations provided in the paper do not offer enough information to support a clear understanding of the process. 4. Regarding multimodal task specification, it is puzzling why RoboCLIP-video (with a different object) yields the best performance instead of RoboCLIP-Edit. One possible interpretation is that the robot is primarily learning push-like behaviors from a video demonstration. However, this raises concerns about the overall effectiveness and usefulness of the edit method in achieving the intended goals.

Questions

1. What is RoboCLIP-Video in Section 4.2? 2. In Section 4.6, how did you use 4 demos to generate one reward for training?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

I appreciate that the author acknowledges the limitation of unstable fine-tuning in their work.

Reviewer QrrK6/10 · confidence 3/52023-07-10

Summary

This paper proposes using the similarity between two embedded spaces, one from real interaction and the other from visual/textual interaction, to develop the rewards.

Strengths

It's a well-written paper. The idea is simple and clearly stated. The proposed method has been sufficiently evaluated on realistic simulation settings.

Weaknesses

It is possible to make the paper more relevant to machine learning as it looks like a robotics paper. The intuition/rationale behind the simple sparse reward structure is not properly explained.

Questions

1) Can authors comment why the paper is more suitable to the NeurIPS readership than CoRL, RSS, ICRA, etc. readership? 2) Can the authors comment why the simple sparse reward of the two embeddings would work? Given that the video demonstrations are not anywhere closer to the real robot interaction, this reviewer is having a hard time understanding how the model works. 3) Looking at Fig 2, it looks like the model is not great at distinguishing "opening" from "closing." Why has the model not learned that concept? 4) What is the cause of variability of rewards in Fig 3? Is it mainly due to how close the video to the real or some other reason? Since the video embedding seems to be significantly more useful, what does text serve? Are the authors trying to convey that even using text (which is easier) is fine if we don't have a video.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

Adequate

Reviewer EmDU2023-08-17

Reply to Authors

Thanks for your rebuttal. I am still not fully convinced. The learned robot's performance is not satisfactory. For instance, in the task entitled "Robot closing green drawer", the video shows the robot moving away from the drawer, while the drawer magically closes on its own. This demonstrates an improper execution of the task. In addition, "Out-of-Domain Videos to Generate Rewards" seems to fail on multiple tasks. The new confusion matrix is still problematic, as the robot is still confused by the text "close drawer" and the video "open drawer".

Authorsrebuttal2023-08-17

Reply

Thanks for getting back to us! 1. Sorry but we are not sure you have seen the updated videos section. We have therefore relabeled the correct section to look at as **Finetuning with One Demonstration (for Reviewer EmDU)**. This is also now the first section on the website after the title. We would have loved to provide a direct link but the rules don't let us :(. Crucially, the "Out-of-Domain Videos to Generate Rewards" section corresponds to using **zero robotic demonstrations**. Therefore failures are expected (the main claim of the paper is that 1 robotic demo is needed, not zero). When provided with **one robotic demonstration, all of the failures are resolved**. In this correct section (please search for **EmDU** on the site and this will take you to the section to view), you will find the failure policy videos on the left column which correspond to the policies learned with **zero** robot demos (the same ones learned from the OOD videos). On the right are the same policies fine-tuned using **1 demo**. As you can see, with one demo, all policies complete the task. This validates the main claim of the paper, **that in all experimental setups, one demo is enough to learn a policy using RoboCLIP**. 2. The reason why the drawer close video looks the way it does is because of rendering. The robot makes contact with the drawer in literally the first few steps of the rollout. Since closing the drawer requires very little force, this interaction is enough to finish the task. In fact, the video uploaded on the site **receives a true reward of 1050** for this seed due to how efficiently it closes the drawer (**baselines get about ~300** and the scripted policy defined in Metaworld gets 481.34). We posit that the reason the robot moves away from the drawer is that the VLM has never seen drawer-opening videos with a robot in them. Thus to maximize the RoboCLIP reward and align the demo to the training data it has seen, it moves away from the drawer ASAP once the required force to close the drawer has been applied. 3. The confusion matrix **is diagonal heavy**, i.e., for each of the tasks, the correct string has the highest correlation with the correct videos. The way to look at the matrix is row-wise. This is because we want to ensure that for a given task (i.e., row corresponding to the videos), the highest entry should correspond to the string describing the task. Thus from the data, there is **no confusion**. We posit that there is some correlation between the remaining pairs because the VLM attends to the "green drawer" in the text and the green drawer object in the video resulting in a correlation. But for each task, **the highest correlation entry corresponds to the correct pair.** Further, the main reason to believe the method works is the performant policies obtained across multiple repeated seeds.

Reviewer Dcaj2023-08-20

Official Comment by Reviewer Dcaj

I would thank the authors for their thoughtful rebuttal. The points raised in the rebuttal have effectively addressed the majority of my concerns. Notably, the revised confusion matrix reduced the noise and showed a more pronounced diagonal distribution. However, I must emphasize that this work maintains an overly direct approach and exists within a landscape containing numerous similar works, which could limit its potential for significant impact. As a result, I'd love to raise my evaluation score to weak accept.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC