CV

Ph.D. Candidate

California Institute of Technology

• Developed MorphMimic, sharing one morphology-conditioned world model and motion-tracking policy across Unitree G1, AgiBot X2 Ultra, and Booster T2. Achieved over 99.7% clip completion in simulation and deployed the shared policy on Unitree G1 for motion tracking and teleoperation.
• Developed CSF, a training-free contextual safety filter for text-conditioned motion generators that grounds natural-language safety rules in reference trajectories enforced through a CBF-QP. Validated on Unitree G1.
• Developed a mobile sensor deployment system using a quadrupedal mobile manipulator for autonomous sensor deployment in the wild.
• Built HANDOFF, a whole-body task-space controller and agentic planning pipeline that enables Unitree G1 to complete natural-language and vision-driven tasks without task-specific data collection.
• Developed PAC-MAN, a perception-aware CBF-RL framework combining real-time vision and control barrier functions for safe whole-body humanoid dodgeball.
• Constructed social-semantic safety frameworks for mobile robots, validated on Unitree G1 and Go2.
• Developed pipelines for data collection, finetuning, and deployment of VLA models for Franka Emika.
• Designed reinforcement learning and hardware deployment pipelines for humanoid whole-body control.
• Developed methods to enforce safety in learned control and planning policies, improving robustness of autonomous systems.
• Created push-recovery and uncertainty-aware locomotion frameworks, validated on humanoid platforms.

Student Research Assistant

University of California, Berkeley

• Conducted research on locomotion, perception, and multi-robot collaboration, with deployments on Unitree A1 and MIT Mini Cheetah.
• Built multi-robot frameworks integrating LiDAR and depth sensing for collaborative load-moving tasks.
• Developed vision-based navigation and obstacle detection stacks in ROS for autonomous quadrupedal navigation.
• Applied reinforcement learning (PPO) and Bayesian optimization to locomotion control, improving stability and adaptability.

Software Engineering Intern

Samsara Inc.

• Delivered production features in a four-person engineering team for fleet management products.
• Built machine learning benchmarking tools to compare models at scale and improve ML engineering workflows.
• Used C++ and Go for backend development and performance optimization.

Ph.D. Candidate, Mechanical Engineering

California Institute of Technology

Robotics, Controls, and Dynamics

• Expected 2027

M.S., Mechanical Engineering

California Institute of Technology

B.S., Electrical Engineering and Computer Sciences

University of California, Berkeley

MorphMimic: World Model and Policy Sharing Across Humanoids

• Shared one morphology-conditioned world model and motion-tracking policy across Unitree G1, AgiBot X2 Ultra, and Booster T2, completing over 99.7% of 128,786 clips on each robot in simulation.
• Deployed the shared policy on Unitree G1 hardware for motion tracking and teleoperation.

Safe Learning with Control Barrier Functions

• Integrated control barrier functions into reinforcement learning training through in-training safety filtering and closed-form discrete-time constraints.
• Validated obstacle avoidance and stair climbing on Unitree G1 without a runtime safety filter; extended the framework to perception-aware humanoid dodgeball.
• Built a training-free contextual safety filter for text-conditioned humanoid motion generators.

Agentic Loco-manipulation for Humanoid Robots

• Developed an agentic planner and a teacher-student framework that distills complementary whole-body controllers for zero-shot loco-manipulation without task-specific demonstrations.
• Deployed on Unitree G1 for table-top sorting, object transport, and bimanual handling.