Robots need not imagine the future at every replan: Bind-WAM routes on language-object binding uncertainty and reaches 98.60% mean success on LIBERO
Synopsis
World Action Models (WAMs) predict future visual states and actions together, but generating a visual future at every replanning step is expensive and sometimes unnecessary. Bind-WAM uses whether the roles named in a language instruction are stably bound to objects in the current view as an interpretable signal: an offline pipeline labels reference mismatch, invisibility and identity drift from complete demonstrations, and a lightweight router of about 6.65M parameters switches a shared WAM between direct action prediction and joint future-action denoising. On LIBERO, Bind-WAM reaches 98.60% mean success while invoking the joint path at only 40.58% of replanning states, cutting average completion time from 15.73 s for Fast-WAM-Joint to 11.71 s.
Interpretation
A robot policy no longer imagines a future frame at every replanning step by default; binding uncertainty between the current observation and the language instruction decides whether joint future-action prediction is invoked. Earlier world action models either removed future imagination entirely at deployment or ran joint generation at every step; AdaWAM selects reasoning modes by execution phase but offers no signal for whether the current state actually needs future prediction. On LIBERO, Bind-WAM reaches 98.60% mean success, above the reproduced Fast-WAM at 97.1% and Fast-WAM-Joint at 98.4%, while cutting average completion time from 15.73 s for Fast-WAM-Joint to 11.71 s, with the joint path triggered at 40.58% of replanning states.
The decision rests on whether the roles a language instruction assigns to objects hold stably in the current view, as when the bowl is the affected object and the plate the relational reference in placing the bowl on the plate. An offline labeling pipeline uses Qwen3-VL-32B to identify task-relevant entities, SAM 3.1 to generate instance masks and propagate them along the trajectory, and DINOv3 ViT-L/16 features pooled inside each mask to measure identity consistency, tagging reference mismatch, invisibility and identity drift as uncertain. These binary labels take only the current RGB observation and language instruction as input to train a router of about 6.65M parameters; on LIBERO and RoboTwin 2.0 the predictor reaches balanced accuracy of 87.59% and 96.75% and AUROC of 0.9653 and 0.9984.
The two inference paths share one set of parameters and differ only in which video tokens the action tokens may attend to. The direct path lets action tokens attend only to the current frame, while the joint path lets them attend to all video tokens and so exploit denoised future visual latents. The router also compares the change in uncertainty between consecutive replanning steps rather than the absolute value alone; on LIBERO, binding-uncertainty routing reaches 98.60% success versus 97.35% for random routing and 97.50% for uniform routing at nearly the same future-prediction ratio, with average completion time falling from 14.58 to 14.64 s to 11.71 s.
Perspective
The demonstrated scope is the LIBERO and RoboTwin 2.0 simulation benchmarks plus three real tasks on a HEXFELLOW ARCHER-D6Y bimanual platform, towel folding, handover-and-place and shape alignment, each evaluated over 11 trials; offline labeling uses only top-head, front or central camera views.
The joint-denoising ratio is non-monotonic on LIBERO: success is 98.15% to 98.40% at roughly 30% to 33%, peaks at 98.60% at 40.58%, and gains nothing when raised further; the thresholds set the fraction of joint-path inference, and the authors list feeding task-outcome feedback into router training and extending to more real tasks and platforms as future directions.