Autonomy
Enables the rover to navigate, perceive, and manipulate its environment without direct human control.
The Autonomy subsystem gives the rover the ability to independently traverse unknown terrain and perform manipulation tasks with the robotic arm. It combines Visual SLAM (VSLAM) for localization and path planning with computer vision for object and marker detection, allowing the rover to complete GNSS-only navigation, AR tag post-finding, and object-retrieval missions autonomously.
We chose a VSLAM based approach over relying solely on GNSS because GPS accuracy alone isn't sufficient for precise final-approach navigation to posts and objects, especially in terrain with signal degradation. VSLAM lets the rover build a local map in real time and correct its position using visual features. For the arm, our only supported autonomy stack involves RL - which is still in development. Using behavior trees and excluding an observation module, we trained a hierarchical reinforcement learning model to dynamically press keys on a keyboard in simulation. Overall, we prioritized modularity between the navigation stack and the manipulation stack so each can be tested and tuned independently before integration.
The subsystem has to operate over unstructured, uneven outdoor Mars-like terrain without reliable GPS lock at every point along the course, using onboard cameras as the primary sensing source. It must reliably detect AR tags, and small ground objects under variable outdoor lighting. The arm's vision-guided keypress task demands sub-centimeter positioning accuracy despite vibration and imperfect stopping position from the drive system.
Tech Specs
- Task completion accuracy
- 98.7%
- Method
- RL with domain randomisation, excluding observation module
