- Rating
- 1
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 2
Read review
Paper summary
This paper proposes a topology-based subgoal discovery method for offline goal-conditioned RL. The authors construct a graph from successful trajectories, learn a geometric representation of the state space, and use persistent homology to identify bottleneck regions that successful trajectories must pass through. These necks are then organized into a hierarchical planning structure and converted into waypoints for low-level policies.
Critical strengths
The authors propose a complete pipeline for neck discovery, including graph construction, geometric representation learning, and topology-based bottleneck detection. The resulting necks can be used to generate a sequence of waypoints, which provides a natural way to decompose long-horizon goal-conditioned tasks into simpler subtasks.
Critical weaknesses
-
The proposed pipeline still relies heavily on human-designed task-specific choices. For example, the authors use different state representations and distance measures for AntMaze and Kitchen, with manually specified choices about which state dimensions are relevant. The entire pipeline is also quite complicated, making it difficult to imagine applying it automatically to arbitrary real-world goal-conditioned tasks without substantial human effort. I think this is a particularly important issue because waypoint discovery can always benefit from additional human priors. In the extreme case, since these benchmarks have fixed environments and relatively simple tasks, one could even manually specify a set of waypoints to guide the policy. It is therefore unclear how much of the performance comes from the proposed discovery method itself, rather than the task-specific design choices.
-
The theoretical analysis is extensive, but its connection to the actual method is not sufficiently clear. The paper devotes roughly 25 pages to theorems and proofs, yet I find it difficult to understand why all these theoretical results are necessary for the proposed approach. I would like the authors to explain more clearly what problem each theorem addresses, which claim in the main paper it supports, and what it tells us about the practical method.
-
The evaluation is limited to relatively simple and well-established benchmarks. AntMaze and FrankaKitchen have been widely used since D4RL (2020), and HIQL, one of the main baselines, was published in 2023. Although the paper also uses the more recent OGBench (2025), the evaluated tasks are still mostly maze navigation and simulated manipulation in fixed environments. Given the paper's broader claims about goal-conditioned control and cross-embodiment transfer, I would like to see evaluations on more challenging and realistic tasks, together with comparisons against stronger recent methods.
Questions for authors
Please see Weaknesses.