Regresar

Slip-aware adaptive path planning for whole-body mobile manipulators using reinforcement learning

Abstract:

This work investigates Reinforcement Learning-based motion planning and obstacle avoidance for a skid-steer mobile manipulator (SSMM), accounting for a unified model of the manipulator arm and mobile base that captures wheel slippage. A slip-aware kinematic model is integrated into the control framework, explicitly accounting for longitudinal and lateral slip to improve pose estimation. These slip dynamics are included in the reward function to discourage unstable trajectories and promote terrain-adaptive behavior. Four RL algorithms - DDPG, PPO, SAC, and TD3 - are implemented and evaluated across simulated scenarios with varying obstacle layouts and target configurations. TD3 achieves the highest task completion rate (83.58%), outperforming SAC (74.76%), DDPG (31.40%), and PPO (7.25%). SAC also produces smoother trajectories, supporting coordinated end-effector movement. Results demonstrate the effectiveness of RL for robust, terrain-aware navigation in SSMMs operating in uncertain and deformable environments.