Event

CW-WAM: A Force-Sensitive Manipulation Model Jointly Predicting Contact Torque

1 reports · 1 sources

Houzhu Ding et al. propose CW-WAM, a world action model that jointly predicts video, motion, and contact torque for force-sensitive manipulation. The paper claims that, compared with a WAM without torque input, CW-WAM raises the success rate of square peg-in-hole insertion with a 1 mm clearance per side from 25% to 70%, and the success rate of power adapter insertion from 40% to 95%. The model uses paired causal attention to structure motion and torque dependencies, and uses a fixed-length torque memory to retain historical interaction information. The predicted torque segments serve as force references for a 20 Hz outer-loop admittance controller, forming a slow-fast control system. The paper claims that after adding the admittance controller, the force-reference RMSE for peg-in-hole insertion drops from 4.8 N to 2.7 N, and for the wiping task from 2.0 N to 1.3 N.