Simulation training

Chubu

Bipedal robot

Chubu is the bipedal robot I’m building. It learns to lift and place its feet from reference motions in MuJoCo, then uses reinforcement learning to improve walking balance.

Side and front three-quarter views of the Chubu simulation model, showing its joints, links, and feet
Chubu in MuJoCo · Two viewing angles
Robot joints
10
Policy update interval
20 ms
Simulator
MuJoCo

Mechanics and motor driver

Mechanical layout shown standing upright, taking a step, and squatting
Mechanical layout · Standing, stepping, and squatting
Top layer of the joint motor driver PCB
Joint motor driver · Top PCB layer

Gait training

Reference gait

A gait clock and reference trajectories guide the leg swing, foot lift, and touchdown.

Reinforcement learning

PPO optimizes the policy. Every 20 ms the policy outputs target angle offsets for all 10 joints. Training accounts for balance, foot slip, and touchdown impact.

Current progress

The mechanical layout, joint driver schematic and PCB layout, and bill of materials are complete. Each leg has three hip joints, a knee joint, and an ankle joint.

The current work is simulation training and motion checks. Physical assembly and hardware testing come next.