Evanston, Illinois, United States
Implemented reinforcement-learning locomotion controllers for modular robots inspired by Reconfigurable Legged Metamachines, training a single universal policy across diverse morphologies by randomizing body configurations, mass/inertia, joint limits, and contact properties in Isaac Lab. Built an end-to-end training pipeline for locomotion primitives (stand-up, walking, turning, recovery) with curriculum + reward shaping (stability, velocity tracking, energy, slip penalties), enabling rapid transfer to unseen configurations without per-robot retuning. Developed a sim-to-real workflow for physical modular legs: actuation calibration (torque/velocity mapping), sensor-noise + latency modeling (IMU/encoders), and domain randomization (friction, backlash, ground heightfields) to reduce deployment failures and improve robustness under hardware variance.
Developed simulation applications integrating real-world datasets (nuScenes, nuPlan) to run both open-loop and closed-loop tests on state-of-the-art (SOTA) end-to-end self-driving algorithms. Integrated dataset pipelines within the simulation framework for open-loop and closed-loop testing on 2,000+ nuScenes scenarios (each 45 seconds), feeding camera images into the end-to-end model and comparing outputs against ground truth to rigorously benchmark performance across diverse driving conditions. Contributed to large-scale training efforts for SOTA models (UniAD, VAD) originally trained on the smaller nuScenes dataset, adapting them to the larger nuPlan dataset to improve model accuracy and reliability. Collaborated with cross-functional teams to facilitate potential deployment of the retrained models on real self-driving vehicles in partnership with industry collaborators.