Residual QP-based Imitation Learning for Cooperative Bimanual Manipulation
Year: 2026
Type: conference-paper
Abstract: Cooperative bimanual manipulation under closed-chain constraints often induces internal forces–interaction forces that do not contribute to object motion but degrade stability and increase actuator load. Conventional imitation learning approaches primarily focus on reproducing demonstrated motions and therefore do not explicitly regulate such physically inconsistent forces during execution. In this paper, we propose Residual QP-based Imitation Learning (Res-IL), a framework that enables force-consistent bimanual manipulation by explicitly minimizing internal forces while preserving task-relevant behavior. Res-IL augments a nominal imitation policy with a residual quadratic program (QP) that computes minimal corrective actions using wrist force/torque (F/T) feedback. To avoid inconsistency between training and deployment, the same residual correction is integrated into the policy fine-tuning loop, enabling the policy to adapt to force-aware action distributions. We validate the proposed method on contact-rich cooperative manipulation tasks in simulation. Experimental results show that Res-IL significantly reduces internal force proxies–by up to 39.0% in force norm and 25.1% in torque norm–while maintaining task success rates comparable to the baseline. These results demonstrate that explicitly incorporating internal force regulation into imitation learning is important for achieving physically consistent and stable cooperative manipulation.
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204195086