Robot preprint reports task success gains from encoding contact and force in actions

Researchers report in a preprint that their ILA method improved robot success rates by 21.0-36.5 percentage points relative to tactile-conditioned direct-action baselines, averaged over four real-world tasks across three policy families. The method represents actions through learned codes that capture how movement, contact and force evolve together. A shared decoder converts those codes into executable actions. The approach supports further training of pretrained robot models and training policies from scratch. For offline improvement, an expert reference model learned from demonstrations guides learning from successful attempts and corrections to failed attempts. The reported gains cover four tasks, limiting the demonstrated scope.