[Thesis Defense] Learning Intelligent Contact for Dynamic Robots

When a robot's foot strikes the ground or its hand presses against an object, interaction forces propagate through the kinematic chain and appear as joint torques and motions, providing a proprioceptive sense of touch. Foundational work in learning-based legged locomotion implicitly processes proprioceptive information to achieve forceful tasks, as presented across the first three chapters of this thesis, which demonstrate systems for highly dynamic running, contact-parameterized walking, and leg-based object manipulation. Subsequently, the core of this thesis develops novel learning-based formulations for controlling contact-force interactions in legged robots. First, we address force regulation through virtual force fields that simulate resistance during training, demonstrating that whole-body force control is achievable through reinforcement learning without dedicated force sensors. Second, we train policies rewarded for estimating physical properties, allowing informative probing behaviors to emerge; when friction estimation is rewarded, the robot learns to scuff its foot while walking, generating informative shear forces. Third, we advance the safety and generalization of humanoid control by reframing compliant whole-body control as an augmented motion-imitation problem, enabling humanoids to yield to external forces while learning from human movements. Together, these contributions show how data-driven control can treat contact as a source of information and opportunity for control rather than a disturbance to be rejected.