BRACE improves how quadcopters reach goals and avoid obstacles in shifting wind

Researchers led by Jaeyoun Choi have presented BRACE (Belief-guided Residual Adaptive Control), a hierarchical reinforcement learning method for reaching a goal while avoiding obstacles under unknown disturbances that change from place to place, in a paper posted on arXiv.

Obstacles redirect wind and terrain changes traction, so disturbances can vary sharply across space, the authors write. BRACE infers a belief about the disturbance from errors in its model's predictions and uses it to condition a high-level policy that adjusts the reference trajectory, combining a nominal term with a residual term guided by the belief.

In three simulated quadcopter settings with suddenly switching wind fields, and on Crazyflie hardware with disturbances from fans, BRACE reached goals and avoided obstacles better than model predictive control and reinforcement learning baselines, the authors report.