Robotics framework evaluation demonstrates precise goal-reaching and safe fault recovery in a 6000 kg mobile robot, highlighting the safety of constrained hierarchical control.
Key Points
To develop and evaluate a robust, hierarchical vision-based goal-reaching architecture that ensures safe deployment and precise tracking on large-scale mobile robots.
Integrated stereo visual localization with a constrained reinforcement learning (RL) planner tailored to mechanical limits and visual smoothness.
Combined a deep neural network feedforward actuator model with logarithmic-barrier-based robust adaptive control (RAC) and a supervisory safety monitor.
Conducted physical validation on a 6000 kg skid-steered robot navigating asphalt and loose-soil terrain, including two injected localization-fault test cases.
Achieved approximately 3–4 cm SLAM-frame final-position root mean square error (RMSE) across test runs.
Demonstrated improved actuator-level command tracking relative to two standard RAC baseline methods under bounded disturbances and slip.
Triggered deterministic braking followed by latched safe-return operation successfully in both injected-fault scenarios.