Overview
Force-sensitive robotic manipulation requires continuous awareness of contact dynamics and external loads—data points that are not explicitly captured by visual observations alone. To address this limitation, researchers developed CW-WAM, a world action model designed to jointly forecast video, motion, and contact wrench. By embedding physical interaction parameters directly into predictive modeling, the architecture generates time-aligned motion-wrench references to improve multi-axis force regulation and task execution.
Research Context
Visual observations provide spatial geometries and scene context but do not explicitly convey the forces and contact loads exerted during physical interactions. When robotic platforms perform tight-tolerance assemblies or surface-following motions, the absence of force and load information often leads to excessive contact forces, unstable interactions, or task failure.
Approach
CW-WAM takes inspiration from hybrid motion-force control principles. The architecture incorporates several dedicated modeling and control components:
- Unified Interaction Module: Formulates both kinematic motion and contact wrench as concurrent modeling targets, outputting time-aligned motion-wrench chunks to serve as control references.
- Pair-Causal Attention Mechanism: Explicitly structures the dependencies between motion trajectories and contact wrenches.
- Long Fixed-Length Wrench Memory: Stores and maintains interaction histories over extended time horizons to inform ongoing predictions.
- Slow-Fast Control System: Couples the core model with a 20-Hz outer-loop admittance controller. The predicted wrench chunks act as reference inputs for the admittance controller, maintaining continuous force feedback and smooth trajectory execution between slower model query cycles.
Findings
Comparative evaluations against standard world action models (WAM) without wrench inputs showed substantial performance gains across multiple manipulation benchmarks:
- Square Peg Insertion: In tests featuring square pegs with 1-mm clearance per side, CW-WAM elevated insertion success rates from 25% to 70%.
- Power Adapter Insertion: Task success increased from 40% to 95%.
- Force Reference Tracking: Implementing the framework with the 20-Hz admittance controller reduced force-reference root-mean-square error (RMSE) from 4.8 N to 2.7 N for peg insertion and from 2.0 N to 1.3 N during surface wiping tasks.
- Stability and Adaptability: The system exhibited smaller force fluctuations during successful executions and improved lifting-height accuracy when manipulating objects of varying loads.