UCLA

Achieving reliable humanoid navigation requires simultaneously reasoning over long planning horizons and satisfying short-horizon dynamic and safety constraints. Classical planners based on visibility graphs, when paired with model predictive control (MPC), can produce efficient collision-free trajectories, yet their effectiveness hinges on carefully hand-tuned parameters and accurate system models. On physical hardware, however, actuation delays, state-estimation errors, and imperfect low-level locomotion tracking routinely cause trajectories to deviate from their nominal paths, producing constraint violations even when the planned path is geometrically sound.This thesis introduces RAVEN (Reinforcement-Adaptive Visibility-Graph Planning with Collision-Free MPC), a hierarchical framework that addresses these limitations by coupling a learning-based global planner with a constrained local controller. Rather than tuning MPC cost weights or substituting the planner with a neural network, RAVEN trains a reinforcement learning meta-policy to adapt the geometric structure of the visibility graph by adjusting per-obstacle inflation radii online. Modifying these radii reshapes the navigable free space and thereby shifts the topology of the global path, allowing the system to proactively compensate for delay and tracking error before they propagate to the controller. A collision-free MPC layer then follows the adapted path while enforcing velocity and obstacle-avoidance constraints over a receding horizon. Training under injected control delays and observation noise encourages the meta-policy to discover inflation strategies that remain robust across varying system conditions, while the underlying geometric planner and constrained optimizer are left intact. Experiments compare RAVEN against a fixed-parameter visibility-graph MPC baseline and a pure end-to-end RL policy in both simulation and hardware deployment on a bipedal humanoid. RAVEN achieves substantially lower obstacle penetration depth under delay than the classical MPC baseline, along with shorter average path length and faster task completion than either baseline, and its simulation behavior transfers reliably to the physical platform. These results suggest that learning to adapt the planning geometry, rather than the controller parameters, is an effective and interpretable strategy for robust humanoid navigation.