- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper studies long-horizon, goal-conditioned planning with a frozen latent world model (LeWM). The authors argue that planning failures worsen as the goal offset grows, and attribute this to three "mismatches": a target mismatch, a search mismatch, and a cost mismatch. They then propose LeWM++, an external planning layer with three components. LatentPathFlow uses flow matching to generate a short latent path and takes its endpoint as the local target. The Action Chunk Prior is a goal-conditioned action-chunk policy trained with IQL/AWR, used to initialize CEM or to inject candidates into it. Min-over-Horizon scores each rollout by its closest distance to the local target. Experiments cover the LeWM Control Suite at multiple goal offsets, visual OGBench, the DINO-WM deformable-object tasks, and real-robot tasks, together with component ablations and sensitivity analyses.
Critical strengths
-
The problem is meaningful, and the framework itself has value. How planning can be improved once the world model is frozen is a question worth studying. Designing planning as a layer outside the world model has a certain degree of generality, and the paper applies the same interface to two world models, LeWM and DINO-WM. The goal-offset analysis on the Control Suite clearly supports one part of the argument: the gap between the full method and the variant without LatentPathFlow grows steadily with goal offset, which shows that local targets are genuinely useful for distant goals.
-
Substantial experimental effort. The evaluation spans several simulated benchmarks and real-robot tasks, with fairly complete ablations and runtime measurements.
Critical weaknesses
-
Much of the experimental work demonstrates that the framework is effective, rather than directly supporting the core claim. The abstract asserts that all three mismatches worsen with goal offset, but the paper does not measure any of them directly; the main evidence is how much performance drops when the corresponding component is removed. In Table 10, only the gap from removing LatentPathFlow grows consistently with goal offset; the gaps from removing the Action Prior or MoH stop growing after H = 50 and even shrink. For the search and cost mismatches, the evidence only shows that the corresponding components help. It does not show that these mismatches exist as described, or that they worsen with distance.
-
The framework design is overly complex, which reduces its contribution. The framework contains several models that must be trained: each environment or dataset requires separately training LatentPathFlow, a value function, two chunk critics, and a policy, and moving to DINO-WM requires retraining all of them in the new latent space. How the Action Prior is used also differs considerably across benchmarks. All of this adds complexity and weakens the generality that an "external planning layer" should provide.
Questions for authors
-
Naming of the method. The name "LeWM++" suggests an improved version of LeWM, yet the LeWM++ planning interface does not modify the world model (which is kept frozen), and the same interface is also applied to DINO-WM (as DINO-WM++). Could the authors explain what "++" refers to, and how the framework relates to LeWM? If the intent is a planning layer independent of any specific world model, the current name may lead readers to take it as a new version of LeWM.
-
Source of baseline numbers. Some baseline results appear to be taken directly from prior papers without attribution: the DINO-WM results in Table 1 match Figure 6 of the LeWM paper, and the DreamerV3 and TD-MPC2 results in Table 3 match the DINO-WM paper. Other baseline results differ noticeably from those reported in the original papers: the LeWM result on PushT in Tables 1 and 10 is substantially lower than in the LeWM paper, and the DINO-WM result on Rope in Table 3 is clearly worse than in the DINO-WM paper. Please clarify which results are cited and which were re-run, explain these discrepancies, and provide comparisons against the originally published results.
-
Implementation of LeWM. Appendix A.2 states that the LeWM used in this paper is a JAX re-implementation with an IMPALA-small encoder, whereas the original uses a ViT encoder. Is the LeWM baseline this same version, and is it representative of the original LeWM? This matters especially on OGBench, where LeWM fails almost completely.
-
Direct execution of the Action Prior. This baseline is reported only on OGBench and DINO-WM, not on the Control Suite or the real robot. Without it, one cannot tell how much of the improvement comes from the planning interface and how much from the offline-learned policy, particularly in settings where most CEM candidates are sampled from the prior.
-
Choice of K_π. The number of prior-sampled candidates differs considerably across benchmarks. How was it chosen, and how sensitive are the OGBench and real-robot results to it?
Minor issues:
- The averages of the full method at H = 50 and H = 75 are inconsistent between Tables 10 and 15 on the one hand and Table 13 on the other.
- The "w/o Action Prior" result in Table 1 differs from the "zero initialization" result in Figure 6(a). Are these the same setting?
- The same ablation is referred to as "w/o LatentPathFlow", "w/o Subgoal", and "w/o Subgoal Path"; the caption of Table 5 is unclear.
- Section 4.6 states that Table 1 and Figure 3 establish the contributions of the three components, but Table 1 has no result without LatentPathFlow and Figure 3 lacks the other two ablations; these results appear only in Table 10.