Overview
I designed and built a physical, standing, and walking bipedal humanoid robot entirely from scratch. The project spans the full robotics stack, bridging mechanical design, custom electronics, physics simulation, reinforcement learning locomotion, and real-time embedded control on physical hardware.
The mechanical architecture was developed in Onshape CAD, optimizing the leg linkages, joint ranges of motion, and mass distribution for dynamic stability while keeping the structure lightweight and rigid. To support power delivery and motor communication, I designed and routed custom PCBs to handle regulated power distribution, sensor bus interfacing, and low-latency communication across all actuators.
For locomotion, I trained neural network control policies using deep reinforcement learning in NVIDIA Isaac Sim and MuJoCo. By introducing domain randomization over joint friction, payload disturbances, and actuator latencies, the learned policies transferred successfully from simulation to physical reality. On the hardware, a ROS 2-based control framework executes real-time state estimation, IMU filtering, and motor command streaming at high frequencies.
Demonstration
The video below demonstrates the physical bipedal robot maintaining dynamic balance, standing, and walking using the learned locomotion policies.