- Rating
- 1
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
The paper proposes Hi-WM, which collects human corrections inside a learned action-conditioned world model instead of on the physical robot. A pretrained policy (Diffusion Policy or π0) runs in closed loop in the world model. An operator takes over by keyboard, leader arm or VR when the rollout looks likely to fail. Cached states allow rollback and branching into multiple corrective continuations. The world model uses the encoder-dynamics-decoder architecture of Wang et al. (2026), trained with added failure and edge-case data. The method is tested on three bimanual YAM tasks: Fold Towel, Push-T and Route Rope. Hi-WM improves real-world success over the base policy and over WM-CL, a baseline that fine-tunes on successful autonomous world-model rollouts with retained data matched. About half of the gain over the base policy remains after this data-matched control (Table 2).
Critical strengths
- The real-robot evaluation covers two policy classes and three tasks, two of them with deformable objects. All methods within a task use the same initial configurations (Table 2, App. B.1).
- WM-CL is matched on retained trajectory count and action steps (B.3). This controls for added data volume, so it isolates human-corrected world-model data from autonomously generated world-model data.
- Edge-case augmentation reduces positioning error between world-model targets and robot execution by more than an order of magnitude (Fig. 6). This is the property that lets corrections made in the world model transfer to the robot.
- The cost projection is explicitly labeled illustrative, not measured (Fig. 7b caption). The convention for counting shared prefixes is disclosed (B.3).
- Gains hold on held-out Push-T scenes with appearance, background and distractor variation (Fig. 8). This suggests the corrections do not overfit to the training scene.
Critical weaknesses
-
The new claim is not examined on its own. The stated observation, that the most valuable supervision lies near the learner's own failure states (Sec. 1), is the premise of DAgger and HG-DAgger (Ross et al., 2011; Kelly et al., 2019). The paper borders on presenting a system rather than a method. The underlying idea is HG-DAgger with the physical robot replaced by a world model, plus state caching. The new claim, that a learned world model can stand in for the robot in the correction loop, is never tested directly. The question that decides whether this works is whether the states the policy reaches in the world model, and the outcomes of human corrections there, match those on the real robot. No measurement addresses it. Two that would:
- comparing failure modes or state distributions of the same policy in the world model and on the robot;
- replaying collected corrections on the robot from reproducible failure states, which is feasible at least for Push-T.
-
Unestablished world-model coverage where the policy lacks it. The approach offloads the policy's out-of-distribution issues onto the world model. It never establishes under what circumstances the world model can be expected to have data and accuracy where the policy does not.
- Failure data is a plausible source of this asymmetry, since it informs dynamics but cannot be imitated. Its source and amount are not reported, however. If it comes from real rollouts of the base policy, robot execution and resets are still required. The saving would then be decoupling operator time from robot time, which is narrower than the claim that collection "requires no execution on the target robot" (Sec. 1).
- For the edge-case data, it is unclear why the policy is not also trained on it.
- World-model fidelity is evaluated only near demonstration states (Fig. 6, Table 1, Sec. 3.2.2). It is never evaluated at the policy's failure states, or for corrective actions that do not appear in the training data.
- Prior-information symmetry: the world model sees real data the base policy does not. This affects the comparison against Base and the cost argument. It does not affect the comparison against WM-CL, which shares the same world model.
-
The additional burden on the human is not characterized.
- Not reported: operator-hours, number of operators, interventions, branches per branch point, collection wall-clock time, world-model inference rate, and the input device used for the reported results.
- Relative to physical correction, the operator works from a generated single low-resolution view with no force feedback. The operator also audits the plausibility of world-model outcomes, since human judgment takes precedence in screening (B.4). Operators may unknowingly exploit optimistic world-model dynamics, and branching increases the number of decisions per failure.
- Without these measurements, it is not possible to characterize whether this burden is worth the gains. One option is success versus cumulative operator-hours for Hi-WM (with world-model data collection counted as a fixed cost), WM-CL, physical HG-DAgger, and additional targeted demonstrations.
- Fig. 7b, an illustrative projection with undisclosed assumptions, appears in the teaser Fig. 1 under "Cost effective scaling." The collection-time analysis in App. D is an accounting identity with none of its terms measured. The scaling result (Fig. 7c) has two increments, for DP only, and no WM-CL curve at matched budgets. That is too few points to show a trend or a scaling advantage over the alternative.
Questions for authors
Questions
- What are the source and amount of the failure and edge-case data used to train the world model? Does collecting it require running the policy on the robot?
- How do Hi-WM corrections compare with physical HG-DAgger corrections at matched operator time?
- What are the measured operator-hours, interventions, branches per branch point, input device and world-model frame rate?
- What fraction of collected corrections succeed when replayed on the real robot from matching initial states?
- Was WM-CL collected over two rounds, like Hi-WM? Is the world model retrained between rounds?
- Deformable dynamics from a single top view are where world-model accuracy for corrective actions seems least likely. What explains gains over WM-CL on Fold Towel and Route Rope that are comparable to Push-T?
- What is the automated rollout assessment? Does post-training use only intervention segments or full trajectories, and how are shared prefixes weighted?
- Is there one world model per task? What checkpoint initializes it, and does π0 start from the public pretrained checkpoint?
- Which backbone is used in Fig. 8, and how does WM-CL perform on the held-out scenes?
Minor and presentation
- Sec. 3.2.1 cites an edge-case proportion ("approximately 30% of the regular-scene data volume") that does not appear in Table 1, which also uses a different denominator.
- WM-CL is used in Sec. 3.1 before it is defined under Baselines.
- Neither Fig. 7a nor Fig. 8 states its number of points or which backbone it uses.
- Trial counts are repeated in the Table 2 caption, Sec. 3.1 and B.1.
- Reference formatting: the 1X World Model entries.