- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
Vision-language-action (VLA) models learn without supervision from future states, while world action models (WAMs) that condition actions on predicted futures add inference cost and let prediction errors propagate into actions. This paper introduces Forward Dynamics Models (FDM), which jointly predict actions and future observations under an asymmetric attention mask: future prediction reads the action expert's hidden states , action prediction never reads future tokens, and the future loss backpropagates into , so the deployed policy runs action-only. FDM is instantiated on , with an added generation expert that regresses one future-image latent, and on Wan2.2-5B, with an added action expert trained jointly with future-video flow matching. The paper evaluates both on LIBERO, RoboTwin 2.0 and five real-robot tasks, and reports an action-to-future ablation, a future-perturbation analysis, gradient alignment and efficiency measurements.
Critical strengths
-
The ablation design is clean. Four variants (action-only, context-only future, stop-gradient, full FDM) separate future supervision, forward conditioning on , and backward supervision through . The stop-gradient control keeps the shared-context gradient paths and detaches only the action interface (App. C.3), so D vs. C isolates the gradient path. The ordering A B C D holds in both backbone families (Table 4).
-
Action-only deployment with real-robot gains. FDM improves over by 13 points and over an architecture-matched WAM by 5 points on five real-robot tasks (Fig. 3). Its inference latency matches (294 vs. 298 ms/chunk), and FDM-VGM is faster than the matched WAM (Table 5).
Critical weaknesses
- Training a policy with an action-conditioned future-prediction objective and dropping that objective at inference is an established design. DriveVLA-W0 (ICLR 2026) trains an action-conditioned world model jointly with the policy and bypasses image generation at inference. UWM (RSS 2025) obtains its policy by setting the future-observation timestep to its maximum. VideoVLA (NeurIPS 2025) already shows that letting video tokens attend to action tokens improves success. What remains new is the specific asymmetric mask over action hidden states (Table 6).
- The closest baseline is missing and reports a higher RoboTwin score. Fast-WAM corresponds to FDM's context-only variant B. With the same Wan2.2-5B backbone, no embodied pretraining and the same RoboTwin protocol, it reports 91.8% (91.88 clean / 91.78 randomized), above FDM-VGM's 89.23% (Table 3). The paper positions FDM against Fast-WAM (Sec. 2) but omits it from Table 3. As a result, the claim that action-to-future supervision beats context-only future supervision is supported only by the real-robot ablation (Table 4).
- LIBERO is saturated and the RoboTwin gains are small. Every method in Table 2 scores between 94.1% and 99.4% on LIBERO. On RoboTwin, FDM-VLM improves over by only 1.45 points and is lower in the clean setting (81.54 vs. 82.74). FDM-VGM is 2.97 points below LingBot-VA (Table 3).
- The "action-grounded supervision" interpretation is not established. The generation expert reads hidden states computed from the noised demonstrated action chunk (Eq. 2; App. A.2), and these hidden states also carry context features. Table 4 shows that supervising this interface helps, but not that the benefit comes from modelling action consequences. An action-swap test or a numerical-action bottleneck would be needed to show this.
- The robustness advantage holds by construction. FDM's zero sensitivity in Fig. 4 follows directly from the attention mask. The WAM's action error rises by only 0.22–5.52% at , and no closed-loop evidence ties WAM failures to prediction errors.
- Minor:
- Table 2 and Sec. 4.1 state "without robot policy pretraining", yet FDM-VLM is initialized from
pi05_base(App. B.1). - Step time throughput in Table 5 implies a batch of 128 for every row, while App. C states 64; please state the timing configuration.
- Table 2 and Sec. 4.1 state "without robot policy pretraining", yet FDM-VLM is initialized from
Questions for authors
- Can you add Fast-WAM to Table 3 and explain why FDM-VGM scores below it with the same backbone and protocol? Does your variant B, trained with your recipe, reach Fast-WAM's reported 91.8%?
- With the observation and instruction fixed, does the predicted future change appropriately when only the action chunk is swapped?
- Which weights are loaded from
pi05_base, and do the real-world comparisons use the same initialization?