← Paper Digest

Bind-WAM: Binding-Uncertainty-Guided Selective Future Imagination in World Action Models

Chenhao Qiu, Shuyi Chen, Songen Gu, Shufeng Nan, Tianyu Wang, Siyu Lin, Yifan Li, Bingwen Zhu, Yanwei Fu

November 2026

TL;DR Bind-WAM lets a world action model imagine future video only when the objects named in the instruction are hard to pin down in the current image, and otherwise predicts actions directly. On a real bimanual robot it averages 66.6% success against 57.6% for always imagining, while imagining at 28.4% of replanning steps.

What they do

World action models (WAMs) predict future video and actions together, but imagining the future at every replanning step is slow and not always useful. Bind-WAM imagines only when language-object binding uncertainty is high, meaning the objects the instruction refers to are hard to identify in the current image.

How it works

  • Offline labels. On demonstrations, Qwen3-VL-32B picks out the manipulated object and the reference objects, SAM 3.1 tracks their masks and DINOv3 features check that each object keeps its identity. A frame counts as uncertain if a reference is mismatched, an object is invisible or its identity drifts.
  • Predictor. A 6.65M-parameter model reads one RGB image and the instruction and predicts this uncertainty. The labeling models are not used at run time.
  • Routing. One shared WAM (Wan2.2-5B video backbone) has two attention masks. On the direct path, actions attend only to the current frame; on the joint path, they also attend to denoised future video. The joint path runs when the uncertainty or its change since the last replan crosses a threshold.

Results

  • LIBERO: 98.6% mean success (Fast-WAM 97.1%, always-imagining Fast-WAM-Joint 98.4%), with imagination at 40.6% of replans. Completion time is 11.71 s, against 15.73 s for Fast-WAM-Joint (25.6% less).
  • RoboTwin 2.0, seven bimanual tasks: 88.3% in clean scenes (AdaWAM 88.4%) and 87.0% in randomized scenes, the highest of the compared methods.
  • Real bimanual robot, 3 tasks, 11 trials each: 66.6% average success, against 57.6% for Fast-WAM-Joint, 42.4% for Fast-WAM and 51.5% for Motus, with imagination at 28.4% of replans.
  • With the same share of joint steps, uncertainty routing beats random (97.35%) and uniform (97.50%) routing and takes about 20% less time.

Limitations

The paper has no limitations section. The real-robot test covers three tasks on one platform with 11 trials each. In simulation, success is only 0.2-0.9 points above always imagining; the main gain is speed. The router learns from perception labels, not from whether imagining actually helped the task, which the authors list as future work.

Figures

From Figure 1: Fast-WAM never imagines the future, Joint-WAM always does, and Bind-WAM imagines only when binding unc...
From Figure 1: Fast-WAM never imagines the future, Joint-WAM always does, and Bind-WAM imagines only when binding uncertainty calls for it. Bars compare success rates, how often each method imagines, and LIBERO completion times.
From Figure 2: Offline labels for reference mismatch, invisibility and identity drift train a lightweight uncertainty...
From Figure 2: Offline labels for reference mismatch, invisibility and identity drift train a lightweight uncertainty predictor. At inference a router picks direct action prediction when uncertainty is low and stable, or joint future-action denoising when the uncertainty or its change crosses a threshold, within one shared WAM.
From Figure 6: Successful real-robot runs of the three bimanual tasks (towel folding, bottle handover into a box, sha...
From Figure 6: Successful real-robot runs of the three bimanual tasks (towel folding, bottle handover into a box, shape alignment), seen from the overhead camera.
From Figure 7: LIBERO success rate and completion time as the share of replanning steps that use the joint path varie...
From Figure 7: LIBERO success rate and completion time as the share of replanning steps that use the joint path varies. Squares mark direct-only and joint-only execution; the star marks the selected setting.
Abstract

World Action Models (WAMs) can improve robot learning by jointly predicting future visual states and actions, but generating a visual future at every replanning step is expensive and sometimes unnecessary. This raises a critical question: when should a robot policy imagine the future? In this paper, we propose Bind-WAM, a binding-uncertainty-guided WAM for selective future imagination. Bind-WAM introduces language-object binding uncertainty, selectively invoking joint future-action generation when the current observation does not support stable task-relevant action generation. An offline labeling pipeline uses complete demonstrations and action context to identify task-relevant objects and detect reference mismatch, invisibility, and identity drift. The resulting pseudo-labels supervise a lightweight router that predicts binding uncertainty from the current observation and language instruction. At inference, a shared WAM policy invokes selective future imagination according to the predicted uncertainty and its change between consecutive replanning steps. Simulation and real-world experiments demonstrate that Bind-WAM can establish binding uncertainty as a strong, interpretable signal for selective future imagination.

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.