Task-Effective Capability-Guided Selection of Arm–Waist Kinematic Subspaces for Humanoid Pick-and-Place Planning

This paper presents a capability-guided method for arm–waist kinematic subspace selection to reduce unnecessary waist motion during local pick-and-place planning for a mobile humanoid. The simulated platform comprises a REEMAN Big Dog mobile base, a custom-designed lift, two seven-degree-of-freedom (DoF) arms, and a two-DoF waist. The proposed method addresses local manipulation after the base position and working height have been set. During simulated manipulation, the base, lift, and right arm remain fixed. The left arm transfers a 40 mm cube between targets in a primary workcell with 18 endpoints, with waist yaw and pitch included in planning when selected. The proposed selector ranks four arm–waist subspaces using endpoint capability information before invoking the motion planner. The motion planner then attempts to generate a complete task-cycle path in the selected subspace. If that attempt fails, the selector activates the next candidate. A held-out evaluation comprised 1728 trials across six methods on 48 task geometries fixed before capability annotation. The proposed method achieved an 80.90% planning success rate. Across 216 task–repetition pairs from 40 tasks on which both the proposed and full-space methods succeeded, pooled mean waist travel was 71.30% lower than that obtained by always planning in the full arm–waist space. This reduction was accompanied by 0.476 additional segment planning calls per trial relative to the full-space method. The simulation results demonstrate that task-specific subspace selection reduces waist travel on the common-success set at the cost of additional planning calls. Validation on physical hardware remains future work.