Data Quality in Imitation Learning

In supervised learning, the question of data quality and curation has been over-shadowed in recent years by increasingly more powerful and expressive models that can ingest internet-scale data. However, in offline learning for robotics, we simply lack internet scale data, and so high quality datasets are a necessity. This is especially true in imitation learning (IL), a sample efficient paradigm for robot learning using expert demonstrations. Policies learned through IL suffer from state distribution shift at test time due to compounding errors in action prediction, which leads to unseen states that the policy cannot recover from. Instead of designing new algorithms to address distribution shift, an alternative perspective is to develop new ways of assessing and curating datasets. There is growing evidence that the same IL algorithms can have substantially different performance across different datasets. This calls for a formalism for defining metrics of "data quality" that can further be leveraged for data curation. In this work, we take the first step toward formalizing data quality for imitation learning through the lens of distribution shift: a high quality dataset encourages the policy to stay in distribution at test time. We propose two fundamental properties that shape the quality of a dataset: i) action divergence: the mismatch between the expert and learned policy at certain states; and ii) transition diversity: the noise present in the system for a given state and action. We investigate the combined effect of these two key properties in imitation learning theoretically, and we empirically analyze models trained on a variety of different data sources. We show that state diversity is not always beneficial, and we demonstrate how action divergence and transition diversity interact in practice.

Paper

References (55)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer AXdc2/10 · confidence 5/52023-06-30

Summary

The paper analyze several theoretical property of the expert demonstration distribution to the performance of the learnt policy in imitation learning. The paper claims that people should provide consistent and transition and state diversity.

Strengths

The paper defines several quantity to clear express the meaning of the properties. The theorems are sound to me.

Weaknesses

The results claimed by the paper are straightforward to me, which makes the contribution less significant. The theoretical results are basically quick derivation of standard machinery of MDP analysis, which are trivial and straightforward. The paper claims that expert should provide consistent demo based on BC analysis. However, the paper does provide experiment of the types of imitation learning methods like transformer architecture to handle multi-modal distribution.

Questions

See weakness.

Rating

2: Strong Reject: For instance, a paper with major technical flaws, and/or poor evaluation, limited impact, poor reproducibility and mostly unaddressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

2 fair

Presentation

2 fair

Contribution

1 poor

Limitations

See weakness.

Reviewer WXvd6/10 · confidence 3/52023-07-06

Summary

This paper considers how to evaluate the quality of a given dataset for an imitation learning task. It is well-known that IL suffers from distribution shift. While much work on IL focuses on improving algorithms, the authors propose that another important approach to making IL more effective is to curate better datasets and demonstrations from experts. Existing approaches to evaluating the quality of datasets focus on state coverage / diversity, trying to ensure that the expert demonstration contains instances of a variety of different states. The authors discuss in some depth how state diversity is an incomplete measure of dataset quality, specifically because it ignores the actions taken in the dataset; the goal of the paper is to find an improved metric for evaluating dataset quality. he paper proposes action diversity and transition diversity. The authors then consider the distribution shift from the expert policy to the learned policy under a given f-divergence (KL divergence by default), as having a large distribution shift can cause problems in IL. They then show a number of properties of their measure and relate it to action diversity and transition diversity , and discuss how these metrics can be applied to improve IL in general (4.3, Implications for Data Curation). They claim that their work suggests that action consistency in the expert demonstrations leads to better outcomes, while state diversty and system noise in the dataset are good or bad depending on context. Finally, they perform experiments on two simple simulationtasks, comparing the effects of adding policy noise to adding system noise in the training set and comparing how demonstration datasets of varying (known) quality correlate with various metrics.

Strengths

The main idea of finding improved metrics for dataset quality, above and beyond state space coverage, is interesting. Furthermore, the dichotomy that their analysis exposes between expanding state space coverage via higher action divergence versus higher transition divergence is an interesting one, showing that not all state space coverage is created equal. The theoretical results linking Q with action divergence are also interesting and well-presented. While there are questions left by the experiments (see Weaknesses and Questions) one extremely interesting finding is that the additon of a small amount of system noise vastly improves the effectiveness of training with policy noise (compare Figure 2 (c)(d) with Figure 2 (i)(j)), especially when the number of demos is low. The effect of this is so strong that it would have been better to focus more on it when discussing the results. With a few minor exceptions, the paper is also very well written. The essential ideas are clearly stated and easy for the reader to absorb.

Weaknesses

This paper would be greatly improved by reducing repetition and adding more experiments. The points made in the Intro are repeated at least 5 times. The paper would be much better if the authors agree to remove this redundancy and add more thorough experiments. he biggest weakness is a key property of the proposed measure (equation (6)), which is tied to a strength of the work ("data quality is heavily tied to the algorithm A that leads to the learned policy \pi_A as well as the expert policy \pi_E"): since it takes the learned policy \pi_A as well as the expert policy \pi_E, actually measuring the quality Q may require running A, which somewhat diminishes the value of the data quality metric since the it may not be usable when collecting the data (note that most common metric, state diversity in the training dataset, can be measured without running A). Even one of the proposed proxies for Q, action divergence, requires having the learned policy to measure against the expert policy. There is an inconsistency in lines 196 and 199-200 with Theorem 4.1: it says the data quality is "lower bounded by the action divergence" but Theorem 4.1 shows that action divergence provides an upper bound to the data quality. Since upper and lower bounds for dataset quality have different implications this is very important to fix. While Figure 2 (i)(j) represents the most interesting finding (in my opinion) of this work, it doesn't escape notice that there is no corresponding Figure 2 (k)(l) showing the same plots for the Square nut task. These should absolutely be included to show whether or not the finding in (i)(j) generalizes to at least one other task (otherwise there is the possibility that (i)(j) is simply a property of one particular setting). The authors do not include a limitations section. Minor weaknesses and typos: In (3), the sum should go for t=1 to T, not H Figure 2 caption: In "X-axis corresponds…" should probably write "injected system noise" twice (rather than "injected Gaussian noise" the first time and "injected system noise" the second time). They're both system noise and both Gaussian, right? Writing it differently may convey the mistaken impression that they might not be both Gaussian, whereas the difference is just when the noise is added. Notation is also confusing, it seems \sigma_s is used both to refer to system noise at evaluation (in the figure itself) and to system noise added in the training set (line 342). Table 1: instead of merely stating that state similarty is the "opposite of state diversity" (line 360), it would be much better to have a more precise definition. Additionally, it is claimed that "state diversity once again is not correlated with performance on the task" (line 370, just above Table 1) but Better/Okay/Worse all seem to have similar state similarity values while Proficient Human (PH) has lower state similarity (i.e. higher state diversity) and has significantly better success rates – doesn't this mean that it is correlated? Claims of no correlation should probably come with the actual correlation number, since it isn't obvious from the table that there really is no correlation. Equation (6) should probably be in a Definition since it's one of the key contributions of the work. 4.3 Implications for Data Curation might be better placed after the experiments section.

Questions

This paper would be greatly improved by reducing repetition and adding more experiments. The points made in the Intro are repeated at least 5 times. The paper would be much better if the authors agree to remove this redundancy and add more thorough experiments. A number of the authors' conclusions (regarding what their results imply about creating and evaluating IL datasets) do not seem to follow obviously from their results and experiments. For instance, they claim in 4.3 (line 294-295): "we posit that expert demonstrators should encourage paths with high system entropy [up to a point]" – this doesn't really seem to follow from their experiments, in which system noise is added to a fixed policy, rather than changing the policy itself. The authors claim to show that action divergence is undesirable and leads to worse outcomes, while transition diversity (in moderation) can improve results; this is demonstrated by adding policy noise during training and comparing the resulting policy to that learned from a dataset with system noise. However, this only shows that action divergence is harmful when it results from noise being artificially added (in which case the expert can be viewed as only partly trying to achieve the task, while also being partly random). What about the effect of action divergence resulting from multi-modality (i.e. when many very different courses of action can achieve the goal in a reasonable way)? It is not at all clear that the proposed measures of data quality will work for multi-modal tasks – which have different solutions, e.g. even simple ones like moving around an obstacle (where you can go to the left or right). It would be interesting to consider how this can be used to guide human expert demonstrations; while I'm not convinced by the discussion in section 4.3, at least without more details (see the first item under Questions), there are possibly other interesting avenues to consider, such as adding noise to the human expert's controls when conducting demonstrations.

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

There is no Limitations section in the paper. The limitations of the work are only very briefly addressed (lines 390-392) and need more expanding; in particular, the weakness mentioned (that the data quality metric proposed is very hard to measure in real datasets) is a serious one and deserves a much broader discussion.

Reviewer 9hHD6/10 · confidence 2/52023-07-09

Summary

This paper studies imitation learning (IL) from a theoretical perspective. It proposes the idea that we should curate the data to improve performance, rather than solely relying on designing more robust algorithms, which is a more common practice in the field. The authors propose data quality for a particular IL algorithm, which is the negative of the divergence between the expert policy and the learned policy using that algorithm. They then show theoretically how two properties, action divergence and transition diversity, are linked to data quality. The authors then study the finite-data regime under a very simplistic setting, and finally show some results of studying their metrics on PMObstacle, Square, and Robomimic domains.

Strengths

The presentation / quality of the writing is top-notch. I really appreciated the clear descriptions throughout the paper, starting from the well-written introduction to the explicit problem setting to the way the authors use English to describe and contextualize all the math. Related work section is comprehensive. I like how it addresses both algorithmic and data robustness. I also really like the framing of Equation 6. The insight that data quality is a function of the learning algorithm is very nice.

Weaknesses

I am unfortunately not well-versed in IL theory, so I cannot comment on the novelty of the work with respect to the rest of the field. Therefore, I will only evaluate the paper on the merits of the contributions stated by the authors. My only major weakness is about the unrealistic conditions underlying some of the theory. * One of the conditions of Lemma 4.2 is that states in the support of the expert policy would have low divergence between actions under the expert policy and the learned policy. The authors say in L226 that with a good learning algorithm, this bounding constant \beta would be small, but I don't see why this would be true in general. For BC, maybe it makes sense, but for other IL algorithms, a "good" algorithm could sacrifice exactly matching the expert at some states, for the sake of getting better interpolation/generalization to states outside of the support of the expert policy. * Theorem 4.3 studies only a very simple setting where the environment dynamics are fixed Gaussian with diagonal variance. I'm not sure how much we can take away from this Theorem as this condition will almost never hold in practice. As a consequence, I find takeaways like that in L254 to be misleading because one cannot draw general conclusions from such a restricted Theorem. Other comments: * One more broad concern I have is that most of the robotics domains studied in the RL literature have deterministic dynamics, including Robomimic (ref: https://arxiv.org/abs/2305.14550) and Mujoco. So how can we obtain transition diversity in such settings? * Theorem 4.1 only makes sense with stochastic policies due to the KL divergence. It's not clear how to apply these findings to deterministic policies, which are quite common (and in fact, are used in the experiments of this paper). * L230, L290: It's unclear what the authors mean when they say the data should maximize system noise. That's a property of the environment dynamics, not the expert policy; we don't have control over it! * It is hard to gain too many insights from Table 1. A simpler explanation for the observed success rates is that as task horizon increases while data size stays fixed, BC simply has less data for each state, leading to worse policies. I'm not really sure why action variance and transition diversity come into play here. Minor things: * The authors use the word "expert" throughout the paper. I would rather say the "behavior policy" or "data generation policy", because the whole point of this paper is that we are trying to curate the dataset, e.g. by introducing more diversity, which means the data may not end up looking like what an expert would do. * It would be good to say in the preliminaries that you're in a continuous state/action space setting. * L73: remove "how" * L167: "\pi_A should match \rho_{\pi_E}(s)" --> I think you mean "\rho_{\pi_A} should match \rho_{\pi_E}"? * L167: "learns only" --> "learns only from" * Lemma 4.2 RHS, subscript of the expectation: \in should be \sim?

Questions

L147: I don't see why it's a "test time" objective. You would be optimizing your policy with this f-divergence loss, so it comes into play during training time, right? Theorem 4.1: What does it mean to have a superscript on D_KL? Is this term the same as the LHS of Lemma 4.2? Do any of the intuitions in this paper transfer to state-only IL, where our dataset doesn't have actions at all? This is becoming an increasingly common setting in the new world of large-scale pre-training (e.g., from human videos).

Rating

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

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central 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

Yes, addressed adequately.

Reviewer 24NK3/10 · confidence 4/52023-07-13

Summary

The paper addresses a key issue in Imitation Learning (IL) called as distribution shift. The paper formalize data quality and data curation for IL. The paper propose two fundamental properties to evaluate/measure data quality as "action divergence" and "transition divergence". The authors investigate these two properties in context of data quality for IL and presented theoretical and empirical results. They conducted an experiment on human/expert demonstration data and provide analysis on data noising and data measuring.

Strengths

- The paper is well-written - The idea presented in the paper is very important for IL

Weaknesses

- The work doesn't seems to fit for a high quality conference like NeurIPS - On measuring the data quality and its effect on success rate, a lot of research is being done on learning from mixed data as its really hard to classify expert vs non-expert data. - The paper seems more like a white paper or can be submitted to AI Safety conferences but not for NeurIPS -

Questions

- addressed in Weakness section

Rating

3: Reject: For instance, a paper with technical flaws, weak evaluation, inadequate reproducibility and incompletely addressed 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

2 fair

Limitations

- Authors are encouraged to explicity mention the limitations of the paper

Reviewer zh6Y7/10 · confidence 4/52023-07-16

Summary

Data quality is an important aspect of offline robotics because there is not a large quantity of data. In particular, in imitation learning, it has been shown that naive training policies on low-quality data can lead to distribution shift issues. Most researchers focused on dealing with this issue using algorithm-centric approaches. Instead, this paper advocates for a data-centric approach for addressing this issue. The data-centric approach modifies the data collection process, where the data collection can be broken down into two categories: action divergence and transition diversity. The authors formalize these two categories of the data collection process and provide empirical evidence of their significance.

Strengths

Formalizing and addressing the data quality issues in imitation learning is challenging. This paper proposed interesting ideas to this challenging problem and has the following strengths: - Clarity: The overall paper is well written, and the problem the authors would like to address is clear. The two categories of data collection are also very clear and make it easy to understand the underlying problem dataset attributes. - Empirical Justification: The author's performance experience using synthetic noise, single and multi-human datasets. All of the experiments performed across these datasets provide insight into various aspects of the data collection discussed in the author's paper. These insights helped enhance the author's discussion of data collection attributes proposed by the authors.

Weaknesses

Although the authors proposed an interesting perspective to address a difficult problem in imitation learning, there exist several weaknesses of the paper: - Lack of novelty: Theorem 4.1 was already proposed and proven in [1] (see theorem 3 appendix). That means the action divergence interpretation was already discussed in the literature. Furthermore, transition diversity was already proposed and discussed to address distribution shift issues in [2]. - Lack of explanation: The authors emphasize that past work only focuses on state coverage while ignoring that action's role. But if the policy providing the demonstrations is an expert, then the states provided from the expert should inherently have low action divergence because this is an expert; and an expert provides the optimal action for a given state. - Lack of citation: The section on distributions shift in IL needs proper citation. [1] Imitation Learning as f-Divergence Minimization by Ke et al. [2] Sequence Model Imitation Learning with Unobserved Contexts

Questions

Below are several questions that I have regarding the proposed idea: - What is the difference between noise in the expert policy and sub-optimal expert? If there is no difference, then is action divergence focuses on suboptimal experts, not experts? - In the system noise section, the authors posit that "the expert demosntrators should encourage paths with high system entropy"; how is this concept different than DAgger[1]? - Why is data quality affected by the algorithm? The algorithm simply uses the data while the expert provides the data. In other words, the algorithms consumes the data while the expert produces the data. - If the expert is optimal and the transition is not deterministic -- that means that when the expert takes action in state s it will, with some probability end up one of many states, for example. Expert data with a noisy transition function means you get more state coverage. This implies that the expert and policy will have more overlap. What is the difference of this idea and the idea discussed in [2] -- which cover the case when the expert and the learner policy have state distribution overlap. [1] A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning Ross et al. [2] Feedback in Imitation Learning: The Three Regimes of Covariate Shift Spencer et al.

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

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

2 fair

Limitations

Yes

Reviewer oFdr5/10 · confidence 2/52023-07-21

Summary

This paper systematically studied the data quality assessment problem in imitation learning, proposing new metrics based on test-time distribution shift.

Strengths

The paper is generally well-organized and well-written. The analysis seems to be in fair detail.

Weaknesses

One of the drawn conclusions seem straightforward to me ("increasing state diversity can often come at the expense of action divergence"). Increasing state diversity will bring more challenges for the learning and thus may cause increased action divergence. The paper lacks practical insights, e.g., how the proposed data quality evaluation metrics can help guide the data collection and thus help the imitation learning is unclear. Currently I dot not see the benefits of introducing such data quality evaluation metrics.

Questions

From Eq.6, the proposed data quality metric is calculated based on the expert policy \pi_E and the learned policy \pi_A by applying an imitation learning algorithm A. If one needs to perform imitation learning first and only after that can get to know how the data quality is, then what is the point of data quality assessment? Can the authors provide some explanations on how the proposed data quality metrics can be used to benefit the imitation learning?

Rating

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

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

3 good

Contribution

3 good

Limitations

Not discussed in the paper.

Reviewer oFdr2023-08-16

I appreciate all the efforts the authors made. I am not familiar with imitation learning, but I am convinced by the authors on their contribution to this field. Thus I changed my score to borderline accept unless other expert reviewers point out any flaws of the paper.

Reviewer 9hHD2023-08-16

Thank you for the detailed author rebuttal and for updating the paper according to reviewer feedback. I am satisfied with the author response, but I maintain my score given my lack of knowledge about the field of imitation learning theory. I would recommend an acceptance if the other, more experienced reviewers do not have a problem with it.

Reviewer WXvd2023-08-16

Thanks to Authors for their response to my review.

I appreciate the Author responses to my review and have also read the other reviews and responses. I appreciate that the Authors have added a Limitations section. After careful consideration I maintain my rating of Weak Accept, but I do feel that Data Quality is an important topic and that the authors should be encouraged to continue working on this topic!

Reviewer zh6Y2023-08-18

- I apologize for the confusion regarding the transition diversity data quality. I read the main comment response in great detail and I agree that the core contribution of your work is to formalize data quality in imitation learning in an effort to guide data curation in practice. - Thank you for your thoughtful response this has answered my question regarding the difference between expert and sub-optimal expert policies. - Thank you for updating the related work to include [1] and [2]. - Thank you for the correction on the difference between System Entropy and DAgger. - Thank you for clarifying the data quality concerns and I have read your main comment response - Thank you for the clarification of state Distribution Overlap compared to the Goldilock’s Regime. I have read and understood your responses to my concerns and your general response. I have no further questions.

Area Chair LhJg2023-08-19

AC Note for reviewer: Please provide more detail

Dear Reviewer 24NK, I think the weaknesses need to be more about the content of the paper than whether the venue is appropriate or not. Can you please directly provide comments about the content of the paper, otherwise it is challenging to factor the review into the decision. Thank you for your service!

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC