Deep Reinforcement Learning for Custom Quadruped Locomotion

Project overview

Area	Details
Platform	Custom quadruped robot
Training	Parallel reinforcement learning in Isaac Gym
Observation	Proprioception only; no external terrain sensing

I developed blind locomotion policies in simulation and transferred them to a custom quadruped. The project focused on producing stable behavior from proprioceptive observations while bridging the gap between parallel simulation and the physical robot.

Policy variants

Policy	Inputs	Terrain	Deployment
Blind flat-terrain walking	Proprioception only	Flat ground	Deployed on custom robot
Blind rough-terrain walking	Proprioception only	Rough terrain	Deployed on custom robot

Result

The blind rough-terrain policy handled 15 cm stairs, dynamic terrain, and complex terrain without falling.

Architecture

The training architecture uses a privileged critic during simulation while keeping the deployed actor dependent on observations available on the robot. This supports learning with richer simulation signals without requiring those signals at runtime.

Demonstration

The project media above shows the learned policy on the physical quadruped and the training architecture behind it.