Projects
Robotics, reinforcement learning, humanoid control, and hands-on mechanical design, from simulation to hardware.
Robotics & Reinforcement Learning
So-Arm Manipulation with RL (PPO) in Isaac Lab
Random motion → precise grasping
Trained a So-Arm robotic manipulator with Proximal Policy Optimization in NVIDIA Isaac Lab. The policy learns coordinated, precise grasping entirely from reward signal. The clip on the left is the fully trained agent running live.
Isaac Lab
PPO
Manipulation
Stage 1 · untrained
Stage 2 · mid-training
Loco-manipulation with RL & AMO Controller in MuJoCo
Walk + pick-and-place, one policy stack
Task-level RL (reach, grasp, lift) layered on a pretrained AMO whole-body controller on a Unitree G1 in MuJoCo. The robot walks to objects and manipulates them, demonstrated with a screwdriver pick-and-place task.
MuJoCo
RL
Unitree G1
AMO
Locomotion + manipulation task
Screwdriver pick-and-place
Whole Body Control & Bipedal Walking
Forward, backward & lateral gait
Lyapunov-guaranteed stability
End-to-end humanoid locomotion stack: WBC formulated as a QP over joint accelerations and reaction forces for real-time balancing torques, plus a DCM planner that decomposes CoM dynamics into stable/unstable components for footstep generation under ZMP constraints.
WBC
DCM Planner
QP Optimization
ZMP
Bullet Physics
