Goal-conditioned Offline Planning from Curious Exploration

Curiosity has established itself as a powerful exploration strategy in deep reinforcement learning. Notably, leveraging expected future novelty as intrinsic motivation has been shown to efficiently generate exploratory trajectories, as well as a robust dynamics model. We consider the challenge of extracting goal-conditioned behavior from the products of such unsupervised exploration techniques, without any additional environment interaction. We find that conventional goal-conditioned reinforcement learning approaches for extracting a value function and policy fall short in this difficult offline setting. By analyzing the geometry of optimal goal-conditioned value functions, we relate this issue to a specific class of estimation artifacts in learned values. In order to mitigate their occurrence, we propose to combine model-based planning over learned value landscapes with a graph-based value aggregation scheme. We show how this combination can correct both local and global artifacts, obtaining significant improvements in zero-shot goal-reaching performance across diverse simulated environments.

Paper

Similar papers

Peer review

Reviewer yLbp6/10 · confidence 4/52023-06-27

Summary

The paper presents a method that using a dataset of rollouts and a dynamic model create a graph like structure in-order to train goal conditioned policies. The nodes of the graph are selected based on it high local value function, and connected using a model based mechanism offline that avoid the myopic behavior of other methods.

Strengths

Solid theoretical analysis of challenges on graph based models that create local and global minima. In-depth comparison with other techniques, and prolific appendix with theoretical support, more results, and value function analysis.

Weaknesses

Even if I see the intense amount of effort on the paper, its hard to piece all the contributions together. I'm mean more in the sense of clarity. One way of addressing this for example could be adding pseudocode of the algorithm. The idea is an improvement of other graph based efforts of the past, as well pointed it out by the authors, Salinov et al , and others. I believe is missing a comparison with the line of work of Lowrey et al 2019 ICRL, "https://sites.google.com/view/polo-mpc""Plan Online, Learn Offline: Efficient Learning and Exploration via Model-Based Control" , Hnsen et al 2022 , Neurips, and related work where a value function analysis it also performed, and the mixed of value function and model based planning is done.

Questions

I would be interesting to see the comparison mentioned in the previous section.

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

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

none

Reviewer nGpW6/10 · confidence 4/52023-06-27

Summary

This paper tackles the problem of performing goal-conditioned control using data collected from curious exploration. They first perform an empirical evaluation of several methods for learning value functions from this offline data, and then identify an issue with estimated value functions called $\mathcal{T}$-local optima, and then a value aggregation scheme that grounds the value function in states observed during exploration, which essentially replaces a long-horizon value function estimate with a combination of shorter-horizon estimates. The paper finds that this aggregation strategy combined with model-predictive control is able to improve performance on tasks including maze solving, pinpad, fetch push, and Franka kitchen.

Strengths

- The paper is well-motivated and clearly written. Each section was easy to follow and the hypotheses and experimental settings very sensible. - The observation of $\mathcal{T}$-local optima as an estimation artifact and a cause of issues in goal-conditioned value function learning is very interesting and to my knowledge, has not been discussed before. While the authors’ empirical analysis in showing that $\mathcal{T}$-local optima indeed are the source of pathologies across multiple environments in the supplementary material is a bit less conclusive than Figure 5 would lead one to believe, I believe that this is a solid contribution that can help to direct future research to look into this further. - The authors’ proposed solution to address these local optima via a combination of model-based planning and a (to my knowledge) novel graph aggregation scheme seems sensible and effective. It achieves improved (on maze_large and pinpad) or similar (on fetch_push and kitchen) performance. - The empirical analysis of actor networks versus model-based planning helps to solidify some observations that have been made in prior work.

Weaknesses

- I would recommend being more transparent with the results for an experiment like Figure 5 for the other environments, where the trend is not as apparent. This would be more helpful for readers to understand how well this hypothesis holds up empirically, and while I understand there is a significant amount of noise, that itself is also useful to know about. - While the manuscript places significant emphasis on the unsupervised exploration/curiosity component of the problem setting, it doesn’t appear that any component of the method is actually specific to the curious exploration setting, nor does it leverage any particular properties of curious exploration. It seems like all of the arguments that the paper makes would be equally valid when using data from a random policy or an expert policy.

Questions

- How does the choice of path length in equation 5 affect performance? - As Equation 5 minimizes over a number of paths composed of learned value estimates, does this result in compounding errors and model exploitation as well? - Following my comment in “weaknesses”, are there specific properties of the curious exploration setting that make this method more effective? Will the performance of this method change significantly if presented with data from a random policy or near-expert policy? ** Update after rebuttal : I have read the authors' rebuttals. They address my concerns reasonably well, and I have chosen to maintain my rating. **

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

4 excellent

Contribution

3 good

Limitations

The authors discuss some of the limitations to their methods, particularly the computational requirements compared to policy learning. I don’t see direct potential negative societal impact for this work, so I think the discussion is sufficient, but would appreciate a further discussion of failure cases or limitations particularly of their graph aggregation method.

Reviewer 7YsN5/10 · confidence 3/52023-06-29

Summary

This paper investigates the issue of mitigating estimation artifacts in the value function of off-policy RL. The authors adopt an unsupervised scheme to collect exploratory data, which is subsequently used to train both off-policy and offline RL algorithms. In their analysis, the authors identify a geometric characteristic known as $\mathcal{T}$-local optima, which represents inaccurate value landscapes. To address this issue, the authors propose a combination of model-based planning and graph-based aggregation, resulting in enhanced performance on long-horizon goal-conditioned tasks.

Strengths

* The finding in this paper that learned value landscapes can hinder achieving long-horizon goals is novel and interesting. * The proposed method, which combines model-based planning with graph-based value aggregation, makes sense and is demonstrated to enhance the performance of vanilla TD3. * The empirical evidence provided in Appendix C through visualization effectively demonstrates the presence of estimation artifacts in the value landscapes.

Weaknesses

* I have concern regarding the performance of the proposed method, as it fails to attain comparable results with previously published works, such as [1][2] on Maze2D, [3][4] on FetchPush, and [5][6] on Kitchen. * Furthermore, the planning method incurs higher training and testing costs than prior offline goal-conditioned works [4][6], yet its performance does not meet the same standards. I recommend that the authors check this issue by conducting a comprehensive comparison, and elucidate the advantages of the proposed method over other offline goal-conditioned RL baselines and model-based planning baselines. * The literature review in this work is inadequate as it fails to cite and discuss important prior offline goal-conditioned works [4][6][7][8]. Notably, the work by [8], which involves the creation of a graph from offline dataset, is particularly relevant. Furthermore, this paper overlooks prior goal-conditioned methods [7][9][10] that employ model-based planning. * The evaluation benchmark appears to be limited, especially two environments are simple point-based tasks. To conduct a comprehensive study, it is essential to incorporate additional benchmarks such as AntMaze in D4RL, as well as the environments FetchSlide, FetchPick, and FetchHandReach used in [4]. * The paper would benefit from an improvement in its presentation and organization. A significant portion of the paper is devoted to discussing value learning, leaving only Section 6.3 for the main experimental results. To address this, I recommend compressing or relocating part of Section 2, Section 3, Section 4, and Section 5, perhaps to the Appendix. Additionally, it would be helpful to move some of the interesting evidence from Appendix C to the main paper. [1] Chen X, Ghadirzadeh A, Yu T, et al. Lapo: Latent-variable advantage-weighted policy optimization for offline reinforcement learning[J]. Advances in Neural Information Processing Systems, 2022, 35: 36902-36913. [2] Janner M, Du Y, Tenenbaum J B, et al. Planning with diffusion for flexible behavior synthesis[J]. arXiv preprint arXiv:2205.09991, 2022. [3] Charlesworth H, Montana G. Plangan: Model-based planning with sparse rewards and multiple goals[J]. Advances in Neural Information Processing Systems, 2020, 33: 8532-8542. [4] Yang R, Lu Y, Li W, et al. Rethinking goal-conditioned supervised learning and its connection to offline rl[J]. arXiv preprint arXiv:2202.04478, 2022. [5] Kostrikov I, Nair A, Levine S. Offline reinforcement learning with implicit q-learning[J]. arXiv preprint arXiv:2110.06169, 2021. [6] Emmons S, Eysenbach B, Kostrikov I, et al. RvS: What is Essential for Offline RL via Supervised Learning?[J]. arXiv preprint arXiv:2112.10751, 2021. [7] Li J, Tang C, Tomizuka M, et al. Hierarchical planning through goal-conditioned offline reinforcement learning[J]. IEEE Robotics and Automation Letters, 2022, 7(4): 10216-10223. [8] Mezghani L, Sukhbaatar S, Bojanowski P, et al. Learning Goal-Conditioned Policies Offline with Self-Supervised Reward Shaping[J]. arXiv preprint arXiv:2301.02099, 2023. [9] Nasiriany S, Pong V, Lin S, et al. Planning with goal-conditioned policies[J]. Advances in Neural Information Processing Systems, 2019, 32. [10] Charlesworth H, Montana G. Plangan: Model-based planning with sparse rewards and multiple goals[J]. Advances in Neural Information Processing Systems, 2020, 33: 8532-8542.

Questions

Overall, I appreciate the authors for identifying the problem and proposing a novel model-based planning method to address it. I would like to raise my score if the authors can address the following concerns mentioned in the above section on weaknesses. * The performance is not comparable to that of prior works in several environments. * It is recommended to compare the results with additional baselines and report the computational costs for both training and testing. * Please consider supplementing the literature review of related works. * It would be beneficial to provide additional evidence from other benchmarks, such as AntMaze or FetchSlide/Pick/HandReach. * Additionally, improving the presentation and organization of the paper would be helpful. * Minor: Line 125 appears to contain a typo where $p_{G}$ is mentioned.

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

4 excellent

Presentation

2 fair

Contribution

3 good

Limitations

This paper highlights one limitation: the method does not address estimation artifacts during training. However, another significant limitation is the negative impact imposed by the computational costs associated with training the ensemble model, graph building, and test-time planning.

Reviewer cdVD6/10 · confidence 3/52023-07-07

Summary

This paper addresses the challenge of extracting goal-conditioned behavior without the need for continued environment interaction. The authors tackle this challenge by leveraging unsupervised exploration to collect diverse experiences. The study explores the suitability of different off-policy actor-critic algorithms and investigates the performance gap between utilizing network-based and model-based planning approaches. To overcome global value estimation artifacts, the authors propose a graph-based value aggregation method that aggregates short-horizon estimates. This technique aims to correct potential inaccuracies in global value estimation. The proposed framework is evaluated using various simulated environments.

Strengths

- Analysis of the performance gap between model-based planning and actor networks is interesting. - Extensive experiments

Weaknesses

One notable weakness of this paper is the limited technical contribution. The two main components proposed, model-based planning and graph-based aggregation, are existing algorithmic components commonly used in the field. As a result, the paper lacks novelty in terms of introducing new techniques or approaches. Furthermore, the proposed framework of using model-based planning and graph-based aggregation does not appear to be specifically tailored to the setup of achieving goals after an unsupervised exploration phase without further interaction with the environment. It would be beneficial to clarify whether certain components or analyses of the proposed framework are specifically relevant to this setup. If not, it might be more appropriate to broaden the scope to a generic goal-conditioned RL setup. Additionally, it would be valuable to see an evaluation of the proposed framework in the context of this broader setup.

Questions

- in section 4.1, why do we need sampling negative goals?

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

The authors adequately addressed the limitations and potential negative societal impact.

Reviewer 7YsN2023-08-11

Thanks for your response

Thank you for providing such a detailed response. I appreciate the inclusion of additional baselines and environments, as they enhance the empirical contribution of this paper. After reviewing the supplemented PDF, I am unclear about why many methods outperform the model-based planning approach in Figure B2. Furthermore, considering the reported computational cost, it appears that the method proposed in this paper requires a substantial amount of inference time (30$\times$ or 300 $\times$ slower than TD3). This could potentially be a major limitation of this work. Nevertheless, since the authors have addressed my other concerns, I am inclined to increase my score to 5.

Authorsrebuttal2023-08-13

We thank the reviewer for the fast response, and we are glad that the inclusions were found to enhance the empirical contributions. We would like to briefly provide some further clarifications. Figure B2 suggests that model-based planning can, on average, improve performance with respect to actor networks, when learning from curious exploration data. We argue that this can be traced back to the ability of look-ahead planning to "escape" value artifacts (see Section 6.1). In general, planning with an (inaccurate) model can however also introduce issues, such as model exploitation (although, in this case the same model is shared by all baselines). Moreover, due to its multi-step optimization, the planner can potentially find distant states with a hallucinated high value that would not be found by one-step greedy maximization (i.e. actor networks). While we found this to be a minor issue in practice, it does appear to occur for certain baselines (e.g. offline model-free algorithms in kitchen, see Fig. B2). Finally, we would like to further comment on runtime. While the inference times reported for our implementation cannot be considered real-time, we are hopeful that this limitation can be overcome in the future. One option for improvement is code optimization, as our codebase was not designed for speed. Another option is hardware scaling. As a proof of concept, we benchmarked our implementation of the algorithms on stronger hardware (i.e. NVIDIA A100). We observed a significant speedup: our method can run a 8Hz and 2Hz, without and with graph-based aggregation respectively. We find this to be promising for two reasons. First, the speedups are significant (2.6x and 6.6x, respectively), and bring the method in line with others relying on model-based planning with learned models [1]. Second, we observe that the slowdown due to graph-based aggregation significantly reduces on better hardware (from 10x to 4x). In fact, the only difference implementation-wise when enabling graph-based aggregation is in the evaluation of the final states of imagined trajectories during planning. When graph-based aggregation is not used, this is done through a single forward pass of the critic. Interestingly, when aggregation is used, this can still be done with a single forward pass, albeit for a much larger batch, as pre-aggregation values must not only be conditioned on the environment's goal, but also on goals embedded in the vertices of the graph. The remaining estimates needed for graph search (e.g. pairwise values on the graph) can be precomputed. This suggests that, with improved capacity for hardware parallelism, the slowdown procured by graph-based aggregation would fade. We are happy to engage in any further discussion or clarification on this and other topics. ### References 1) Thodoroff et al. Benchmarking Real-Time Reinforcement Learning, NeurIPS 2021 Pre-registration Workshop

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC