When to Trust Imagination: Adaptive Action Execution for World Action Models

Abstract: World Action Models (WAMs) have recently emerged as a promising paradigm for robotic manipulation by jointly predicting future visual observations and future actions. However, current WAMs typically execute a fixed number of predicted actions after each model inference, leaving the robot blind to whether the imagined future remains consistent with the actual physical rollout. In this work, we formulate adaptive WAM execution as a future-reality verification problem: the robot should execute longer when the WAM-predicted future remains reliable, and replan earlier when reality deviates from imagination. To this end, we propose Future Forward Dynamics Causal Attention (FFDC), a lightweight verifier that jointly reasons over predicted future actions, predicted visual dynamics, real observations, and language instructions to estimate whether the remaining action rollout can still be trusted. FFDC enables adaptive action chunk sizes as an emergent consequence of prediction-observation consistency, preserving the efficiency of long-horizon execution while restoring responsiveness in contact-rich or difficult phases. We further introduce Mixture-of-Horizon Training to improve long-horizon trajectory coverage for adaptive execution. Experiments on the RoboTwin benchmark and in the real world demonstrate that our method achieves a strong robustness-efficiency trade-off: on RoboTwin, it reduces WAM forward passes by 69.10% and execution time by 34.02%, while improving success rate by 2.54% over the short-chunk baseline; in real-world experiments, it improves success rate by 35%.

• The paper introduces FFDC-WAM, a framework that adapts action execution by verifying model-generated predictions against real-world observations.
• The paper quantifies a 69.10% reduction in model forward passes and an improvement in real-world success rates from 45% to 80% in pick-and-place tasks.
• The paper confirms that leveraging co-predicted sensory sequences for self-verification fosters timely replanning and robust, efficient robotic control.

Adaptive Trust in World Action Models: A Formal Analysis of FFDC-WAM

Introduction

The paper "When to Trust Imagination: Adaptive Action Execution for World Action Models" (2605.06222) addresses a foundational limitation of existing World Action Models (WAMs) for robotic manipulation: fixed-length action chunk execution renders robots unaware of divergence between model-generated imagination and real-world outcomes during open-loop rollouts. The study posits adaptive execution as a future-reality verification challenge: the system should execute further only as long as the WAM’s predictions align with observed reality, demanding timely replanning upon detecting inconsistencies.

Problem Formulation and Prior Work

Standard WAMs, building on recent advances in video-action co-prediction and large-scale pretraining, generate both future observations and action chunks conditioned on current sensory input and an instruction token. This improves transfer, generalization, and task flexibility compared to single-modality (VLA) models. However, continual execution of a fixed action horizon is often suboptimal: short chunks sacrifice efficiency by over-querying the WAM during stable phases; long chunks yield high contact failures in stochastic or contact-rich manipulation due to unseen deviations from the model's imagination.

Prior solutions in adaptive execution—such as entropy- or uncertainty-based chunk adjustment (Jing et al., 24 Nov 2025, Liang et al., 5 Apr 2026, Wang et al., 24 Feb 2026)—typically attend only to policy-side action uncertainty, failing to leverage the core property of WAMs: their capacity to explicitly encode imagined future sensory sequences coupled with predicted actions. This work exploits that self-consistency prior for robust adaptive control.

The FFDC-WAM Framework

The proposed system, FFDC-WAM, integrates a standard WAM (Motus backbone (Bi et al., 15 Dec 2025)) with a lightweight, high-frequency verifier module—Future Forward Dynamics Causal Attention (FFDC). After each WAM inference produces a segment of predicted actions $\hat{A}_{t+1:t+H}$ and latent visual tokens $\hat{O}_{t+1:t+H}$, FFDC cross-examines these predicted futures against real observations, planned actions, and language instructions.

Figure 1: FFDC enables the system to condition execution adaptively on consistency between WAM-imagined futures and real rollouts, outperforming fixed-size chunking in robustness-efficiency trade-off.

Instead of a static rollout, FFDC scores an execution-confidence value $e_t \in [0, 1]$ at each step, evaluating whether the imagined horizon remains trustworthy. Execution continues if $e_t \geq \tau$ and triggers replanning otherwise. The emergent effective chunk lengths thus scale with the temporal persistence of model–reality consistency.

Verifier Architecture

The FFDC verifier is a transformer-based attention module using a causally-structured Boolean mask that ensures temporally aligned interaction between predicted actions, predicted visual tokens, current observation, and language embedding.

• Input sequence: $$X_t = [L, \hat{O}_{t_p}, O_t, \hat{O}_{t_f}, \hat{A}_t, \text{[CLS]}]$$, where $L$ are language tokens, $\hat{O}_{t_p}$ are past predicted visual tokens, $O_t$ is the current real observation, $\hat{O}_{t_f}$ are future predicted tokens, and $\hat{A}_t$ is the action segment.
• Efficient execution: WAM-generated context is cached after each inference; FFDC operates as a frequent, lightweight verifier using only the latest real observation per step without redundant large model forward passes.

  Figure 2: Schematic of FFDC-WAM, illustrating module flow and temporally structured causal attention at each check step.

The model employs a tightly-local attention window and strictly causal masking to efficiently enforce cross-modal, temporally correct credit assignment between visual and action tokens.

Training and Dataset Construction

The WAM is trained via mixture-of-horizon sampling for robust long-horizon prediction. The verifier $\hat{O}_{t+1:t+H}$0 is trained as a binary classifier with positive examples from successful rollouts/demonstrations and negatives from failed rollouts plus diversified action-sequence corruptions (including temporal swaps, gripper flips, late-stage noise, and tail scaling).

Experimental Results

Robustness-Efficiency Trade-off in Simulation

On the RoboTwin benchmark (Chen et al., 22 Jun 2025), FFDC-WAM demonstrates improvement in both efficiency and robustness compared to both short- and long-chunk baselines. Notably, it:

• Reduces model forward passes by 69.10% and task completion time by 34.02% compared to fixed short-chunk (Base-Motus) while marginally improving average success rate.
• Outperforms fixed long-chunk methods on "hard" tasks (i.e., those with $\hat{O}_{t+1:t+H}$1 for Base-Motus) by 22.2 percentage points on Rand.hard settings, without sacrificing efficiency on easier tasks.
• Dynamically allocates WAM calls—a clear manifestation of difficulty-aware execution (see Figure 3).

  Figure 3: FFDC-WAM executes long action chunks in predictable phases and triggers timely replanning in difficult task segments, mitigating failures endemic to open-loop fixed-chunk execution.

Real-World Transfer

FFDC-WAM’s capacity to detect model-reality divergence online generalizes to real-world Astribot S1 deployment. Success rate rises from 45% to 80% across two pick-and-place tasks relative to fixed chunk execution. The slightly higher average execution time and WAM calls are justified by increased robustness, as replanning is triggered to correct for noise, errors, and contact uncertainty not observed in simulation.

Figure 4: Real-world examples where FFDC-WAM alternates between execution and online replanning, correcting for observed drift unavailable to open-loop policies.

Ablation Study

Ablations on RoboTwin hard tasks confirm that all four verifier inputs (predicted actions, predicted visuals, real observation, and language) contribute to robust execution. Removing predicted visual tokens has the strongest negative impact, reducing success rates by 4.8 percentage points, highlighting that future observation prediction is the most critical self-verification signal. Real observations and action tokens are also indispensable for high-confidence, low-overhead verification.

Theoretical and Practical Implications

The FFDC-WAM framework suggests that effective online deployment of video-action models requires explicit, temporally structured mechanisms for continual future–reality alignment. The study demonstrates that:

• Self-verification using co-predicted world futures is both computationally feasible and beneficial, enabling dynamic horizon selection unattainable through static chunking or action-only uncertainty proxies.
• The approach improves downstream task performance and reliability in both simulated and physical environments, establishing a new paradigm for integrating model-based imagination into closed-loop adaptive control.

Such modular future verifiers may generalize to other multi-modal predictive policy architectures, as the core principle—explicitly bridging self-consistency between model-generated futures and reality—applies to any setting where models can simulate multi-step environmental trajectories.

Limitations and Future Directions

Key limitations include reliance on a binary supervision signal for training the verifier and the use of a fixed acceptance threshold. Model performance may degrade when facing out-of-distribution mistakes not represented in the synthetic corruption or failed rollout set. Extension to richer continuous-valued or multi-class confidence estimation, threshold learning (potentially with learnable trade-offs), and integration of richer failure data are proposed directions for enhancing verifier realism, adaptability, and robustness. Expanding verification to more complex, multi-agent, or dexterous scenarios remains a fertile research avenue.

Conclusion

This work formalizes adaptive execution in WAMs as a future–reality verification task and advances the field with a causally aligned, lightweight verification module—FFDC. Experimental results substantiate strong gains in both robustness and efficiency, with the practical conclusion that continual self-verification against imagined futures is an essential attribute for future robust world-model-based control systems.