Domain Adaptive Imitation Learning with Visual Observation

In this paper, we consider domain-adaptive imitation learning with visual observation, where an agent in a target domain learns to perform a task by observing expert demonstrations in a source domain. Domain adaptive imitation learning arises in practical scenarios where a robot, receiving visual sensory data, needs to mimic movements by visually observing other robots from different angles or observing robots of different shapes. To overcome the domain shift in cross-domain imitation learning with visual observation, we propose a novel framework for extracting domain-independent behavioral features from input observations that can be used to train the learner, based on dual feature extraction and image reconstruction. Empirical results demonstrate that our approach outperforms previous algorithms for imitation learning from visual observation with domain shift.

Paper

Similar papers

Peer review

Reviewer V2Bv5/10 · confidence 3/52023-06-30

Summary

This paper aims to solve the domain-adaptive imitation learning problem, where the demonstrations are collected in a source domain, and the agent needs to learn policies in a target domain. Beyond that, the tasks are with high-dimensional visual observations. The authors have proposed a new framework based on dual feature extraction and image reconstruction to achieve behavior feature extraction and reward generation.

Strengths

This paper aims to solve a very practical problem, since in real-world problems, we can hardly collect expert demonstrations in the target learning environment. Instead, collecting expert demo in a similar source domain is much easier.

Weaknesses

The framework in Figure 3 is too much complex. The authors are encouraged to clarify their contributions in this paper and simplify the parts in the existing literature. Previous works study the problem of domain adaptive imitation learning, and this work aims to solve the same problem with visual observations. The proposed approach augments the previous work with an encoder to deal with the high-dimensional visual observations. This idea seems a little straightforward.

Questions

The experiment section is weak. The proposed approach is barely evaluated in limited simple tasks. If the authors could show that the proposed approach can work in more challenging domains, such as humanoid running, the quality of this paper could be greatly improved.

Rating

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

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

2 fair

Limitations

This paper has not discussed the limitations in the main body of this submission.

Reviewer 3tvh7/10 · confidence 3/52023-07-04

Summary

Imitation learning from visual observations for domain adaptation. Given expert and non expert demonstrations in a source domain, and non expert demonstrations in a target domain, learn a reward and policy in the target domain via image consistency and domain and behavior feature consistency.

Strengths

Domain adaptation in imitation learning remains an important problem. The paper tackles this challenging problem assuming access to visual observations but not to actions, other proxy tasks or expert demonstrations in the target environment. These assumptions allow for better generality. The problem statement, setup and method are clearly presented and the quality is high.

Weaknesses

- The paper empirically demonstrates shifts in color, number of joints that can move, and action space. What are the general assumptions about the relation between the source and target domains? (There is some similarity, otherwise transfer would not be possible..) - The papers uses visual observations, which is a more general setup than states. However, the proposed domains are quite simple (very few agents/objects in each, very simple backgrounds, etc.). How is the method expected to scale to more complex environments?

Questions

In addition to the weaknesses above, - How do the target policies compare to the source policies? E.g. in RE2 to RE3, does the policy learn to use all 3 links or does it apply a strategy with 2 links? Would suggest providing videos for qualitative analysis. - In Figure 10(a), why is it desired that the source and target expert behavior features overlap? Shouldn't they require different features for different domains? (E.g. number of links.) - Is there an explanation for the performance gap with GWIL (in Figures 5-9)? A state-based method should have more information about the environment. - Please state the state/action space for every environment (for completeness and clarity). - Suggest defining formally what good/bad actions mean. - Line 22: suggest replacing 'should' with 'can'. - Typos in lines: 55, 130, 194, 195, 355.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed 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

Please add to Section G the assumption re access to the target environment for training $\pi_{\theta}$ and in what cases will D3IL fail to learn a successful target policy.

Reviewer szho6/10 · confidence 4/52023-07-06

Summary

The paper attempts to solve the problem of domain shift in Imitation Learning with image observation. The author proposes D3IL architecture which is based on dual feature extraction (behavior and domain specific features from input) and dual cycle consistency (image level and feature level cycle consistency). The paper demonstrates effectiveness of the proposed method compared to baselines Third Person IL (TPIL), Disentan GAIL over different Gym based tasks (e.x. Reacher, Pendulum etc.) and DeepMind Control Suite (DMCS) tasks (Maze based tasks).

Strengths

- The problem addressed in an interesting problem for Imitation Learning. - Good thorough evaluation different types of experiments. - Strong experiments results. - The paper is well written.

Weaknesses

- The problem is well motivated from real robotics based task, but no real robotics based task is demonstrated in the paper. It would be interesting to see the application on sim-to-real based tasks. - It would be good to discuss work similar TCN work [1]. - The extensions to other IL methods like IRL is a little unclear. - Line 90-93 are a little unclear. [1] Sermanet, Pierre, Corey Lynch, Yevgen Chebotar, Jasmine Hsu, Eric Jang, Stefan Schaal, Sergey Levine, and Google Brain. "Time-contrastive networks: Self-supervised learning from video." In 2018 IEEE international conference on robotics and automation (ICRA), pp. 1134-1141. IEEE, 2018.

Questions

- Have you tried interchanging source and target domains? - The relations to source -> expert and target -> learner is a little confusing. Maybe if you can represent in a better way that would be really good.

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

3 good

Presentation

3 good

Contribution

3 good

Limitations

- most of them are mentioned in the paper

Authorsrebuttal2023-08-20

Thank you again for your valuable feedback to review our paper. We kindly mention that the discussion period is about to end. We sincerely hope that our responses and additional experiments successfully address the reviewers' concerns and enhance the clarity of our work. If the reviewers have remaining comments and suggestions, we will be grateful to share them with us.

Reviewer 33qr5/10 · confidence 4/52023-07-25

Summary

This work introduces a method of learning a visual behavior encoding for reinforcement learning tasks, D3IL. From a TPIL behavior feature encoder baseline, D3IL introduces dual feature extraction (including both behavior and domain feature encoders) and dual cycle-consistency (cycle consistency on image and feature levels). From inputs of unpaired target and source images, D3IL translates across domains and then cycles back to the original domain, via a pair of generator networks which take in source images and domain targets. Further, image reconstruction consistency, feature vector similarity, and feature vector consistency is added for further loss and stability. The domain and behavior feature encoders and target generator from above are taken and used to train a reward generating discriminator, $D_rew$, which distinguishes between expert and learner behavior (rather than expert and non-expert data used for training the previous models). This is used alongside the target behavior encoder to train SAC RL policies for the given tasks. D3IL is tested on many domain transfer tasks with variations in visual effects, degree-of-freedom, internal dynamics, and robot embodiment - D3IL demonstrates superior performance on the target domain compared to baselines of TPIL, …. It is further demonstrated on an AntUMaze task where direct RL learning in the target domain is difficult, but possible with D3IL. Ablation studies show that the contributions of each of the consistency losses across feature and images, and cycle consistency, are all important and contribute to the final performance.

Strengths

#### Originality The work builds on top of TPIL to build a larger framework for cross-domain imitation / feature encoding, by incorporating dual feature extraction and cycle consistency, and training a further reward-generating discriminator. #### Quality The experiments are detailed, covering four categories of domain change and with ablation studies on components of the loss. The experiments are run over multiple seeds and the method is compared against several baselines (TPIL, DisetanGAIL, GWIL) to present a good picture of the overall learning and domain transfer capability. There are further many more detailed experiments in the appendix to which present good coverage. #### Clarity The work is generally well-written and the description of D3IL which draws from premises of prior TPIL work is understandable. The explanations of losses and model composition are detailed and the diagrams help with understanding the proposed methodology. The experiments are easy to follow. Further task explanation and experimental methodology is well explained in the appendix. #### Significance The method seems broadly applicable across tasks and domain transfer goals in simulation. The experiment on AntUMaze is particularly motivating as towards the practical applications of the method, demonstrating that a task which could not be solved directly could be improved by training from an imitation agent learned on a simpler task - in a sort of curriculum learning.

Weaknesses

#### Quality The dual feature extraction, in particular the domain encoder added from TPIL, is motivated on lines 135-136 as to “check that the obtained feature well contains the required information without information loss, based on image reconstruction and consistency check.” However, it’s still unclear why specifically it is necessary to pass observations as input to domain encoder, as the work assumes that the policy knows what environment it is operating in. Would an ideal domain encoder be able to output a one-hot vector indicating its domain to minimize all the domain-encoder relevant losses? As there are separate generators for each domain (source/target), should it be clear without any domain conditioning what the “style” of each generator output should be? For example, in CycleGAN and its successors, the domain encoding is not required. An ablation study without using a domain encoder, but a static method of encoding the domain or skipping the domain encoding entirely may be helpful. #### Clarity Unfortunately many of the experimental results are left to to the appendix including all of Sections 6.2 and 6.5, and reading the appendix seems necessary to have a full understanding of the model performance. The method is somewhat hard to initially grasp from reading the text of Section 5 as there are many similar variables in the notation which differ by one letter or modifier, with many such modifiers. However, explanations are generally helped by Figures 1 and 3. The limitations are also only mentioned in the appendix. Misc: - Line 191: is another form of consistency [that] guide[s] the model so that … - Line 195: scheme significantly improve[s] - Line 211: we need [the] expert behavior feature #### Significance The practical applications of the method seem best suited for cases where it is hard to train an expert in the target environment directly via RL, but having access to expert data on other tasks as observation-only would help. Many of these cases would be covered by real world scenarios, which is out-of-scope of the current work. Even on the AntUMaze problem, the learning process did not necessarily have to be done via adversarial learning and visual-only observations, as the full expert state would be available. From the ablation in Table 1, it appears that the large majority of the performance improvement comes from image-level consistency and the feature-level consistency was only of minor gain, especially compared to the uncertainty/error. As mentioned in the limitations in the appendix, the paper also introduces quite the complex process with significantly many more loss terms and two pipeline stages compared to prior work, which may make it harder to apply. Adversarial learning techniques generally require more careful tuning than supervised methods, and this work introduces two adversarially-learned stages.

Questions

Some discussion on the weakness-quality section would be most helpful

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

3 good

Presentation

3 good

Contribution

3 good

Limitations

Some consideration of the limitations is explained in the weaknesses-significance section.

Authorsrebuttal2023-08-20

Thank you again for your valuable feedback to review our paper. We kindly mention that the discussion period is about to end. We sincerely hope that our responses and additional experiments successfully address the reviewers' concerns and enhance the clarity of our work. If the reviewers have remaining comments and suggestions, we will be grateful to share them with us.

Reviewer V2Bv2023-08-17

Response

Thank you for the response. I acknowledge that I have read the rebuttal.

Authorsrebuttal2023-08-20

Response to the Comment

Thank you again for your valuable time and effort to review our paper. We appreciate the reviewer's insights and suggestions to improve the quality of our work.

Reviewer 3tvh2023-08-17

Response to rebuttal

Thank you for the response, it has addressed and clarified some issues regarding the results. My score remains the same since it is still not clear how to generally characterize what is the required relationship between source and target domains to achieve generalization, or how well this method can scale to more real-world domains.

Authorsrebuttal2023-08-20

Response to the Comment

We sincerely thank you again for your valuable comments and suggestions. We appreciate the reviewers' time and effort to review our paper and help improve the quality of our manuscript.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC