Résumé
Ensuring safe physical interaction between torque-controlled manipulators and humans is essential for deployingrobots in everyday environments. Model Predictive Control(MPC) has emerged as a suitable framework thanks to itscapacity to handle hard constraints, provide strong guar-antees and zero-shot adaptability through predictive reason-ing. However, Gradient-Based MPC (GB-MPC) solvers havedemonstrated limited performance for collision avoidance incomplex environments. Sampling-based approaches such asModel Predictive Path Integral (MPPI) control offer an alter-native via stochastic rollouts, but enforcing safety via additivepenalties is inherently fragile, as it provides no formal con-straint satisfaction guarantees. We propose a collision avoidanceframework called COSMIK-MPPI combining MPPI with thethe toolbox for human motion estimation RT-COSMIK andthe Constraints-as-Terminations transcription, which enforcessafety by treating constraint violations as terminal events,without relying on large penalty terms or explicit humanmotion prediction. The proposed approach is evaluated againststate-of-the-art GB-MPC and vanilla MPPI in simulation andon a real manipulator arm. Results show that COSMIK-MPPI achieves a 100% task success rate with a constantcomputation time (22 ms) largely outperforming GB-MPC.In simulated infeasible scenarios, COSMIK-MPPI consistentlygenerates collision-free trajectories, contrary to vanilla MPPI.These properties enabled safe execution of complex real-worldhuman–robot interaction tasks in shared workspaces usingan affordable markerless human motion estimator, demon-strating a robust, compliant and practical solution for pre-dictive collision avoidance (cf. results showcased in https://exquisite-parfait-ffa925.netlify.app).