Autonomous navigation of planetary robotic rovers through sandy, rock-strewn terrain requires that onboard algorithms provide accurate models of the rover kinematics, sensors, and environment in order to localize the robot in its map accurately. Rover slip presents a major challenge in such an environment since formulating analytic slip models a priori using terramechanics can be difficult. This thesis develops a slip-adaptive navigation algorithm that can model slip empirically online as the rover traverses the environment. The slip-adaptive algorithm essentially augments the rover kinematic model with terms that approximate the unmodelled slip. These slip terms are estimated by a multi-layer feedforward neural network (MLFN), which uses an extended Kalman filter (EKF) to estimate the network weights while it also provides the rover state estimates. Slip is learned and thus detected autonomously, eliminating the need for laborious human commands uploaded once per Martian sol. The Mars Exploration Rovers have already demonstrated the benefits of implementing more autonomy into these systems, including shortened traverse times, increased scientific exploration, and prolonged rover survivability.
展开▼
