Large-model-assisted coordinated motion planning and stability control for humanoid robots
TL;DR
The results indicate that LM-CoMotion provides an interpretable bridge from natural-language motion commands to physically constrained humanoid joint plans, while real-robot validation with wrist orientation, force, IMU, and contact measurements remains future work.
Abstract
Humanoid robots are expected to transform high-level human demonstration commands into coordinated, executable, and stable whole-body motions, but this transformation remains difficult when semantic task parsing, multi-joint trajectory generation, inverse kinematics, and stability verification are treated as separated steps. This paper presents LM-CoMotion, a large-model-assisted coordinated motion-planning framework for humanoid robots. The framework first uses a large language model as a semantic parser to extract action intent, participating joints, motion duration, coordination relations, and safety constraints from a natural-language command. The parsed task is then converted into deterministic motion primitives, including quintic body-yaw planning, mirrored circular arm trajectories, shoulder-elbow inverse kinematics, neutral wrist roll/pitch/yaw constraints, hip-yaw and hip-roll compensation, and center-of-mass stability checking. A representative APMCM humanoid-robot task is used as the case study, in which the robot turns its body left by 45deg while both arms perform opposite circular motions. The experimental section is built only on the original modeling outputs, including joint-angle trajectories, arm-coordination curves, hip-compensation results, and CoM displacement ranges, without adding fabricated hardware data or new images. Compared with a non-boundary-constrained linear baseline and a cubic baseline, the quintic trajectory satisfies both zero-velocity and zero-acceleration boundary conditions, providing smoother start-stop behavior for demonstration motion. In response to dynamic-feasibility concerns, the revised framework further distinguishes static support-polygon containment from hardware-level dynamic feasibility and introduces a ZMP/angular-momentum consistency condition as a necessary validation extension. The results indicate that LM-CoMotion provides an interpretable bridge from natural-language motion commands to physically constrained humanoid joint plans, while real-robot validation with wrist orientation, force, IMU, and contact measurements remains future work.