CQM: Curriculum Reinforcement Learning with a Quantized World Model

Recent curriculum Reinforcement Learning (RL) has shown notable progress in solving complex tasks by proposing sequences of surrogate tasks. However, the previous approaches often face challenges when they generate curriculum goals in a high-dimensional space. Thus, they usually rely on manually specified goal spaces. To alleviate this limitation and improve the scalability of the curriculum, we propose a novel curriculum method that automatically defines the semantic goal space which contains vital information for the curriculum process, and suggests curriculum goals over it. To define the semantic goal space, our method discretizes continuous observations via vector quantized-variational autoencoders (VQ-VAE) and restores the temporal relations between the discretized observations by a graph. Concurrently, ours suggests uncertainty and temporal distance-aware curriculum goals that converges to the final goals over the automatically composed goal space. We demonstrate that the proposed method allows efficient explorations in an uninformed environment with raw goal examples only. Also, ours outperforms the state-of-the-art curriculum RL methods on data efficiency and performance, in various goal-reaching tasks even with ego-centric visual inputs.

Paper

References (50)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer vggh6/10 · confidence 5/52023-07-04

Summary

- The proposed method addresses the challenge of learning a curriculum in goal-conditioned policies, which is a significant problem. Previous research on curriculum in goal-conditioned policies has often overlooked the importance of learning the underlying semantic goal space. This paper builds upon a recently proposed method that utilizes VQVAE to learn the semantic goal space and extends it to incorporate the learning of a curriculum. By selecting a sequence of sub-goals, this curriculum-based approach aids in achieving the final goal. - The paper compares the proposed method to many different SOTA baselines.

Strengths

- The paper is very well-written. - The paper effectively compares the proposed method with various baselines, providing a thorough analysis. - Additionally, the paper conducts ablations to examine the impact of different design decisions made in this research.

Weaknesses

- It will be valuable to assess the effectiveness of the proposed method on both manipulation tasks and more intricate navigation tasks like 3D environments (ex. Habitat, AI2-Thor etc) - It seems restrictive to the reviewer to use single-code representations with VQ-VAE so would be interesting to see how the results change if using multiple codes like in DGRL. (The reviewer understands this is mentioned as future work, but this seems a severe limitation if the proposed method works only with single code and not multiple codes). - A minor comment: To enhance clarity, it is recommended to explicitly mention in the introduction that prior research, such as DGRL [1], has proposed and emphasised the significance of learning a semantic space for goal specification. It was only later in the paper that I realised vector quantisation had been proposed in previous work as an abstraction for goal-conditioned RL. [1] Discrete Factorial Representations as an Abstraction for Goal Conditioned Reinforcement Learning, https://arxiv.org/abs/2211.00247

Questions

- I'm curious about how the performance varies by varying the number of codes in VQ-Dictionary. - It will also be useful to study if the performance of the proposed method can be improved by using self-supervised objectives for learning visual representations as done in DGRL.

Rating

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

3 good

Presentation

3 good

Contribution

3 good

Limitations

Refer to Weaknesses.

Reviewer eib87/10 · confidence 4/52023-07-05

Summary

The method works as follows. Graphs are built by (1) quantizing visual observations to create a goal space & (2) creating temporal relations over goal space vectors. Curriculum goals towards a "user"-specified goal are made using this graph. A VQ-VAE is used to create the goal space where goals are decodings of one of k trainable embedding vectors. The authors also study an ablation where landmarks are decoded from the nearest neighbor goal encodings over replay buffer observations. Graphs are derived by connecting edges based on a metric that exploits having rewards by -1 at all states except for the goal state. This is motivated by wanting edges to capture the geodesic between nodes. A curriculum for exploring the graph is created by selecting nodes with high uncertainty (or low count via count-based methods) and high temporal distance. Curriculum goals are goals that maximize this sum of this uncertainty and geodesic distance away from the initial state. The method uses Dikstra's algorithm to generate plans towards goals. They compare against many baselines in the literature and consistently find a higher success rate. They also visualize the curriculum goals produced by their method.

Strengths

originality: The method is novel to the best of my knowledge. quality: The paper has high quality. The experiments are very though and the authors also do an analysis. clarity: The experiments are relatively clear. significance: The paper seems to produce better performance than many methods in the literature with similar assumptions (e.g. exploiting diikstra's algorithm).

Weaknesses

The paper could strongly benefit from some figures that describe the details of the method. The method is relatively complicated (though not more complicated than comparable methods in the literature). Due its complexity, it's hard to understand how the method works from figure 1. I'm somewhat familiar with this literature so I think understand their method section was easier but other readers may find it more challenging. Another figure would strongly improve the clarity of this paper. The baseline description method describes how the baselines work but it does not describe what comparing against each baseline (or sets of them) tells us about CQM. Without this, while the results are promising, it's hard to get a takeaway message. For example, OUTPACE also proposes an uncertainty and temporal distance aware curriculum. Comparing against OUTPACE tells us....? It might be the importance of jointly considering uncertainty and temporal distance but that isn't clear in the text. I recommend explicating this a bit more. I think it's fine if baseline methods are grouped together for describing the point of the comparison. Given the complexity of this method, it would strongly benefit from an algorithm describing, how is data collected, how is the goal space learned, how is the graph constructed, etc. In particular, which of these are done in tandem and which are done in sequence is not clear. Overall, this method seems promising, the motivation is clear, and the results are also promising. I think the biggest drawback is clarity right now. It's not that easy to understand the method. Right now I lean towards accept but I strongly encourage the authors to add figures/explicit algorithm/both for the methods section. This would likely push my score to a 7.

Questions

Across experiments, does the agent encode first-person observations, top-down observations over a map, or both? This seems important to how you quantize the observations. If top-down, this seems like a strong assumption. Can you justify it? Methods like SFL use first-person observations. At least some of the environments that you study permit first-person observations (e.g. AntMaze). I know you do only ego-centric for PointNMave-Viz but it's not clear for the other experiments. What assumptions does this method make over the reward function?

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

2 fair

Contribution

3 good

Limitations

This limitations section is fine.

Reviewer uT8t8/10 · confidence 4/52023-07-06

Summary

This paper proposes a curriculum RL approach using a VQ-VAE to learn a goal space, and then construct a graph with the VQ-VAE codes as nodes, and a temporal distance estimate of the Q-value as weighted edges. The curriculum is then constructed by doing frontier-based exploration on this graph, by sampling goals based on their visitation count and temporal distance. In addition, the graph can be used for planning intermediate goals as waypoints for the agent. The paper is well written, and it has proper benchmarks and ablation study. The authors use the term "world model" in their paper, but I would argue they don't really train a world model, rather a representation of the observation space. I would omit the use of "world model" in a further revision of the manuscript.

Strengths

The paper proposes a novel curriculum RL method, which learns a goal-space on high-dimensional pixel observations. Turning the VQ-VAE codebook into a graph, weighted by estimated temporal distance from the Q-value is a powerful idea for both curriculum learning as well as planning and long-range credit assignment.

Weaknesses

- The authors use the term "world model", however what they present is a discrete latent representation of the observation space. A world model would entail to e.g. also equip it with an action-conditioned forward dynamics model, as typically used in model-based RL. I would not call this a world model.

Questions

- How to get a distribution over the goal, encoding a set of exemplar goal observations and counting the bins of the VQ-VAE they get encoded in? - In the Planning section the authors use state s_0 from S, although the previous MDP description uses observation o_0 from O. I assume these are the same or have you shifted to a POMDP setup? - In the AntUMaze, the performance seems to deteriorate after 600k steps, i.e. the curriculum goals seem to get farther from the goal. Any insight on what is happening there? - Since the goal space is trained as a VQ-VAE on individual observations, I think it will break when the environment is ambiguous, i.e. if the agent would have to traverse rooms, but 2 rooms would look identical, these observations would map onto the same goal and the graph might collapse?

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and 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

3 good

Limitations

- I think one limitation that is currently not touched upon is that this approach requires that the environment is not ambiguous (i.e. you don't see the same observations at different places in the environment).

Reviewer paJb5/10 · confidence 4/52023-07-07

Summary

This paper introduces Curriculum RL with Quantized World Model (CQM), a novel approach that leverages a VQ-VAE to create a discretized goal space and constructs a graph structure over it. CQM further proposes a curriculum strategy based on uncertainty and temporal distance to guide the learning process. The authors evaluate the effectiveness of CQM through experiments conducted on variants of PointMazes and AntMazes, which serve as benchmarks for Hierarchical Reinforcement Learning.

Strengths

- This paper is well-written and easy to follow up - Extensive experiments

Weaknesses

One significant weakness of this paper is the lack of clarity regarding why using the representation from VQ-VAE is suitable for graph-building. It is crucial to consider the temporal distances between nodes to adequately cover the visited state space, especially given the limited number of nodes. The proposed goal representation learning scheme by VQ-VAE does not appear to take into account these temporal distances, raising questions about the efficacy of the proposed representation in creating a meaningful semantic goal space. It would greatly benefit the paper to provide a more thorough explanation and justification for the use of VQ-VAE in graph construction and its ability to capture temporal relationships. Furthermore, the proposed goal representation learning scheme should be compared to prior work on representation learning, such as NORL [1] or LESSON [2]. A comprehensive comparison would help establish the novelty and effectiveness of the proposed approach in relation to existing methods. Additionally, it remains unclear how the proposed goal representation learning scheme outperforms or differs from the simple approach of utilizing farthest point sampling from the replay buffer, which warrants further investigation and comparison. Moreover, the paper introduces new hyperparameters (\alpha, \beta, and \kappa) for curriculum goal generation. However, it is unclear how these hyperparameters were determined and how the performance varies when these hyperparameters are varied. Additionally, it would be valuable to compare the costs associated with hyperparameter search between CQM and the baseline methods. [1] Ofir Nachum, Shixiang Gu, Honglak Lee, Sergey Levine, “Near-Optimal Representation Learning for Hierarchical Reinforcement Learning”, ICLR 2019. [2] Siyuan Li, Lulu Zheng, Jianhao Wang, Chongjie Zhang, “Learning Subgoal Representations with Slow Dynamics”, ICLR 2021

Questions

- How does performance vary when \alpha and \beta are changed? Could you provide insights into the selection process for these hyperparameters? Is it possible to set them automatically? - Regarding planning over the graph, what happens if an agent is unable to directly reach a specific node w_{i}? Could the system handle a scenario where, after TemporalDist(\phi(w_{i-1}), \phi(w_{i})), the agent conditions its behavior on reaching w_{i+1} instead? - Could you please explain how the landmarks were sampled from the replay buffer for the "CQM Landmark from Replay Buff" depicted in Figure 7? What criteria or process were used to select the landmarks from the replay buffer? Minor: - It appears that there are two different symbols used for \beta, one in Equation 2 and another in \alpha = 1 / \max (\beta + \kappa D_{\text{KL}} (p_{g}^{f} || p_{ag}), 1). I would recommend using distinct symbols to avoid confusion. - In Figure 7, the y-label representing the distance is partially hidden.

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

2 fair

Limitations

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

Reviewer G8ip5/10 · confidence 3/52023-07-17

Summary

The paper proposed a new curriculum reinforcement learning method, CQM, that uses VQ-VAE to learn a quantized goal space, constructs a graph on the quantized goals to propose curriculum goals by distance, and learns a goal-conditional policy.

Strengths

- The proposed method pioneers in "auto" curriculum RL that learns the goal space and proposes goals all by itself. - The paper is clearly written and easy to follow. - Detailed studies on various environments are presented to demonstrate its effectiveness. - Code is provided for reproduction.

Weaknesses

- $\hat{o}_t$ in Eq. 2 should be $o_t$. - [Vector Quantized Models for Planning](https://arxiv.org/pdf/2106.04615.pdf) has used VQ-VAE in RL. They did not explicitly generate curricula but it's very similar to the proposed method. The authors should compare and explain the difference. - $Q(l_i, a, l_j)$ in Eq. 3 is not explained. I guess it's an expectation over the replay buffer? - The proposed mazes are long but they mostly don't have branched dead ends, which contain unseen states but do not lead to the actual goal. [This work](https://openreview.net/pdf?id=U4r9JNyNZ7) and its experiments in more complicated mazes should be compared to.

Questions

- Eq. 6 does not have a weight factor to balance the two terms. How are the two terms distributed in practice? Will one dominate the other?

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

The limitations have been adequately addressed. > After rebuttal Some of my concerns have been addressed. I appreciate the authors' efforts, but the maze in either their submission or rebuttal doesn't match the complexity of the related work I requested for comparison. Thus I'm holding my score.

Reviewer uT8t2023-08-11

I thank the authors for their detailed responses to my questions, and I appreciate the additional results they provided.

Authorsrebuttal2023-08-13

Thank you for your response

Thank you for replying to our response. We appreciate again your valuable suggestions and all your efforts during the review process.

Reviewer paJb2023-08-14

My concern has been addressed

Thank you for the response that addresses my concern. Please add the experimental results and discussion into the final version. I would like to raise the score.

Authorsrebuttal2023-08-15

Thank you for your response

Thank you for replying to our response. We are happy to hear that our response addressed your concern. Following your suggestion, we will discuss this in the final version. We appreciate again for your valuable suggestions and all your efforts during the review process.

Reviewer eib82023-08-15

Reviewer response

Thank you for your response. I am satisfied with the rebuttal. I have raised my score by 1.

Authorsrebuttal2023-08-16

Thank you for your response

Thank you for replying to our response. We are happy to hear that our response addressed your questions. We appreciate again for your valuable suggestions and all your efforts during the review process.

Authorsrebuttal2023-08-21

Dear Reviewer G8ip

We hope this message finds you well. We appreciate the time you have taken to review our work and consider the points we raised in our rebuttal. We hope that our response has provided a more comprehensive understanding of our research and its potential contributions to the field. Please let us know if you need any further clarification. We appreciate again your valuable suggestions and all your efforts during the review process. Thank you for your consideration. CQM Authors.

Authorsrebuttal2023-08-21

Thank you for replying to our response. We appreciate again your valuable suggestions and all your efforts during the review process.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC