Researched learning-based control for agile legged locomotion, specifically enabling a Unitree Go1 quadruped to autonomously dribble a soccer ball through dynamic environments and score goals. The work combined reinforcement learning in simulation with sim-to-real transfer to produce robust, task-driven behavior on physical hardware.

Contributed to end-to-end system development — from environment and task design through policy training to hardware deployment and evaluation.

Impact:

• Implemented actor-critic PPO policies in Isaac Gym and Isaac Lab with separate locomotion and ball-manipulation networks for coordinated gait control and goal-directed dribbling

• Engineered domain-randomized environments with parameterized contact dynamics, friction, terrain variation, and external perturbations to maximize transfer robustness

• Designed multi-objective reward functions balancing ball possession, goal progress, heading alignment, and gait regularity; iterated on shaping terms to eliminate reward hacking

• Profiled and mitigated sim-to-real degradation from actuation latency, state estimator noise, and rigid-body contact model mismatch

• Hardware deployment currently in progress!