ForceWAM splits contact learning in two: demos set the goal, play data teach force response, reaching 95% on one real-robot task
Synopsis
In contact-rich manipulation, expert demonstrations cover successful paths but say little about how force responds to different actions, while task-agnostic play records broader contact dynamics without a task goal. ForceWAM gives the task-compatible motion and six-axis force/torque future to a force world model trained on demonstrations, and gives the question of which actions realize that future to an inverse dynamics model trained on both demonstrations and play; at deployment, fresh wrench feedback corrects actions between world-model updates. Across force-regulated sliding, box flipping with and without disturbances, and ring placement, the full system reaches 85%, 95%, 85%, and 90% success, above the listed baselines.
Interpretation
Contact-rich manipulation is split into two separately learnable questions: what motion and contact evolution a successful interaction should show, and which robot actions can realize it. Earlier end-to-end force-aware policies coupled which transition advances the task with which action realizes it inside one task-supervised mapping, relying on successful demonstrations alone. On four real-robot conditions, force-regulated sliding, box flipping, disturbed box flipping, and ring placement, the full system reaches 85%, 95%, 85%, and 90% success, above every external baseline in the table; the task-demonstration-only variant also beats each external baseline.
The force world model generates future wrist video and future six-axis force/torque trajectories from successful demonstrations, and the inverse dynamics model outputs actions conditioned on the current observation plus a visual-wrench future. Inverse-dynamics training windows come directly from the recorded visual-wrench futures and associated actions in demonstration and play segments, without running the world model first to build training inputs. Play data are added in steps from zero through intermediate subsets to all episodes per task family with the world model and optimizer updates held fixed; sliding rises from 40% to 85%, flipping from 75% to 95%, disturbed flipping from 65% to 85%, and ring placement from 75% to 90%.
At deployment the world model refreshes its reference every six actions, and between refreshes the inverse dynamics model re-denoises the same cached noise with a growing measured wrench prefix, executing only the current action. A visual future used only as a static planning target does not specify how contact force should evolve along that path; here the generated force future also serves as a contact reference. Among four future-conditioning variants, joint visual-wrench conditioning has the highest success in all four conditions; removing perturbation augmentation of recorded futures lowers flipping, disturbed flipping, and ring placement by several points each and sliding by more.
Perspective
The results come from real-robot experiments on a single Flexiv Rizon 4S 7-DoF collaborative arm with an end-effector six-axis force/torque sensor, covering force-regulated sliding, box flipping, disturbed box flipping, and ring placement, with a number of trials per condition; success means task completion without a safety stop or human intervention.
The play-data gain is measured with the world model and optimizer updates held fixed and only the inverse dynamics model's training data changed; end-to-end models given the same play data instead lose success, so the benefit depends on this decoupled structure. The world model's predicted visual detail and peak wrench values deviate from measurements, while the main interaction stages are preserved.