“Rapid Task and Model Adaptation for Legged Robots”

Dept: Mechanical & Aerospace Engineering

Chair: Christian Hubicki, Ph.D.

Abstract

The successful deployment of dynamic robotic systems relies heavily on heuristic hand-tuning, such as objective weight calibration in optimal control or reward shaping in deep reinforcement learning. This dissertation argues that manual tuning merely treats a symptom of a fundamental problem: model mismatch. This mismatch occurs behaviorally when rigid objectives fail to adapt to real-world priorities, and physically when simulated dynamics fail to capture true hardware properties. To address these challenges, this work introduces a computational framework built on in situ task adaptation and physical system parameter identification. To overcome task rigidity, a trajectory optimization planner actively encodes learned failure models to dynamically reprioritize objectives and synthesize control strategies in real time. To resolve physical discrepancies, a system identification pipeline leverages differentiable simulation to regress unmodeled actuator parameters directly from hardware data. Validated across monopod hoppers, micro-aerial drones, and quadrupedal robots, these methods demonstrate quantifiable improvements in real-time task adaptation and sim-to-real transfer. This dissertation establishes a principled foundation for deploying agile autonomous systems that systematically align their underlying models with physical reality.