Obstacle avoidance learning for a wheeled lunar rover based on its local relative pose

A learning approach for the lunar rover's obstacle avoidance is presented. The nonholonomic dynamics of the rover BH-II is modeled. The obstacle is avoided by following the sub-goals sequence in the vision range. The sub-goal is the intersection point of the obstacle's normal line and the maximum vision circle, and its distance to the obstacle is a given value that assure the rover of a safe moving area without collisions. We choose to learn the control trajectory instead of solve the dynamics ODE equations to overcome the heavy computation task. In order to ensure the strategy's robustness, the relative position vectors between the rover, the sub-goals, also and the obstacle are chosen to compose the discrete states. In the learning experiment the translational velocity is set at a constant low speed, so only the rotational torque as the control to be learnt. Considering the relationship among the front or the rear wheels for fulfilling the nonholonomic constrains, six wheels' steering torques are reduced to two. Finally we use the learnt control trajectory to drive the rover to move to the target in the simulation environment with obstacles. Experiment results show that the learnt strategy is adaptive to the changing environment.