UC Berkeley

The deployment of mobile robots in unstructured, real-world environments demands robust solutions for both perception and control. Conventional approaches often rely on precise, expensive sensors and rigid control structures, which can be brittle under noise and constraints. This dissertation explores a shift toward learning-based frameworks to enable robust, efficient, and self-reliant autonomy across aerial and legged platforms. To address the challenge of aerial navigation in global positioning system-denied environments, we first introduce a real-time, image-based geo-localization framework. By utilizing a database of rendered synthetic views from satellite imagery and topography, this two-stage pipeline performs global and local descriptor matching to recover the 6-degree-of-freedom vehicle pose. The system achieves onboard inference at approximately 1.1Hz with a position root-mean-square error of 2.8m, demonstrating robustness to camera pitch variations, lens flare, and motion blur. Transitioning to legged platforms, we address the limitations of stiffness in position-based control by proposing an end-to-end learning torque control framework for quadrupedal robots. This reinforcement learning (RL) approach outputs joint torques directly at high frequency, bypassing conventional low-level proportional-derivative controllers. Validated on a quadrupedal robot, the framework demonstrates superior tracking accuracy and dynamic compliance, enabling the robot to recover from significant external disturbances and traverse compliant terrains such as air mattresses and pebbles. To bridge the gap between control and perception, we present a modular legged state estimation framework. In contrast to concurrent training or teacher-student distillation, which can hinder modularity and tracking performance, we decouple training by optimizing an RL-based estimator against a fixed controller within a unified environment. This approach ensures effective sim-to-real transfer and achieves superior base linear velocity command tracking compared to concurrent baselines. Finally, we propose a cross-modality distillation pipeline to achieve cost-effective perceptive locomotion. We distill the geometric reasoning of an elevation map-based teacher policy into a student policy that relies solely on a monocular color camera using a diffusion transformer architecture. The resulting image-based policy enables a humanoid robot to traverse complex terrain, such as stairs, matching the performance of the teacher without requiring explicit depth sensors or map fusion. Collectively, these frameworks support robot autonomy by integrating learning-based perception and control.