Robotics & Reinforcement Learning
So-Arm Training using RL (PPO) in Isaac Lab
Trained a So-Arm robotic manipulator using Proximal Policy Optimization (PPO) in NVIDIA Isaac Lab. The agent progresses from random movements to precise, coordinated grasping.
Stage 1: Initial (untrained)
Stage 2: Mid-training
Stage 3: Fully trained
Loco-manipulation using RL & AMO Controller in MuJoCo
Combined locomotion and manipulation on a quadruped robot using RL with a pretrained AMO controller. The robot walks to objects and manipulates them — demonstrated with a screwdriver pick-and-place task.
Locomotion + manipulation task
Screwdriver task
Whole Body Control & Bipedal Walking
End-to-end humanoid locomotion system combining Whole Body Control, DCM-based trajectory planning, and multi-directional gait generation. WBC is formulated as a QP minimizing joint accelerations and reaction forces subject to full-body dynamics constraints, providing real-time torque commands for balancing. The DCM planner decomposes CoM dynamics into stable/unstable components for footstep trajectory generation with ZMP constraints. Together they enable forward, backward, and lateral walking with Lyapunov-guaranteed stability.
