ThinkWAM: Rethinking Latent Futures for Action Refinement
TL;DR ThinkWAM proposes an action chunk, uses that proposal to correct its predicted future latents, then regenerates the chunk from the corrected futures with the same action head. It reaches 84.76% on LIBERO-Plus without OOD fine-tuning and 64.4% mean success across three out-of-distribution real-robot conditions, against 56.7% for π0.5.
What they do
World-action models predict future states alongside actions, but usually the predicted future does not depend on the action being proposed, so it cannot separate a good proposal from a flawed one. ThinkWAM adds a proposal-rethink-refine loop: a first action pass proposes a full chunk, that proposal corrects the predicted future latents, and a second pass of the same action head generates the chunk that is executed.
How it works
- Backbone. A frozen V-JEPA 2.1 ViT-L encoder, a Qwen2.5-0.5B language backbone and a flow-matching diffusion-transformer action head (ActionDiT), about 1.5B parameters in total.
- Near-fine, far-coarse prediction. A temporal JEPA predictor outputs future V-JEPA latents at three horizons (t+7, t+15, t+31 on LIBERO) with 288, 144 and 72 tokens.
- Rethink. A residual model reads the first-pass chunk and the current visual tokens and corrects each horizon’s latent; the corrected latents condition the second pass. Both passes use four Euler steps.
- Affordance-guided memory. A decoder trained on contact heatmaps weights the visual tokens and adds a residual to memory slots built from the current observation and recently executed actions.
Results
- Simulation: 98.70% on LIBERO, 84.76% on LIBERO-Plus without OOD fine-tuning, 85.75%/65.85% on RoboTwin 2.0 Clean/Random, and 65.17% on RoboCasa24.
- Real Galaxea A1Z, trained only on single-object pick-and-place and tested with two targets, color distractors and purple light (30 trials each): 64.4% mean success, against 56.7% for π0.5 and 37.8% for FastWAM, at 101.2 ms per chunk.
- Ablation on LIBERO-Plus: executing the proposal instead of the refined chunk drops success from 84.76% to 80.51%; two passes without latent correction give 81.59%.
Limitations
Every chunk needs two action passes, and whether more rounds of feedback pay off is untested. The residual is supervised with demonstrated futures, so the model does not predict what executing its own proposal would cause. The real-robot test covers one pick-and-place task on one arm.
Videos
Figures



Abstract
World-action models (WAMs) couple action generation with predictions of future states. An initial action chunk may be flawed. A future prediction that is not conditioned on the chunk is identical for every proposal and thus carries no proposal-specific signal, limiting its usefulness for refinement before execution. We introduce ThinkWAM, a latent WAM that establishes bidirectional feedback before execution through a proposal-rethink-refine loop: the policy first proposes a complete action chunk, the proposal corrects multi-horizon future latents through a learned residual, and the corrected latents are injected into a second pass of the same action head, which generates the executable chunk. An affordance-guided observation-action memory and a near-fine, far-coarse multi-horizon prediction scheme, both motivated by predictive coding, support this loop, and all prediction stays in latent space. ThinkWAM achieves 98.70% success on LIBERO, 84.76% on LIBERO-Plus without OOD fine-tuning, 85.75%/65.85% on RoboTwin 2.0 Clean/Random, and 65.17% on RoboCasa24. ThinkWAM achieves 64.4% mean success across three real-world OOD conditions. The mean two-pass inference latency per action chunk is 153.48 ms on LIBERO, 164.61 ms on RoboTwin, and 101.2 ms on the Galaxea A1Z. Ablations show complementary gains from rethinking and affordance guidance. Diagnostics indicate that proposal-conditioned corrections change both predicted futures and action trajectories, showing that the corrected future representations improve the final action.
Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.
