Back to projects
Project 01 | Newton
Quadruped reinforcement learning capstone

Newton

A capstone project where we built a low-cost quadruped robot and trained it to walk.

Newton project image
01 / 06

Note

Have you ever thought to yourself "Lets try building a quadruped robot with 12 degrees of freedom?" I had not either, but surround yourself with people who do and you might end up helping bring one to life.

Context

I worked on the reinforcement-learning side of the project while helping keep the project moving across research, simulation, and hardware work. We initally started with Isaac Sim but moved to Genesis since the first setup was annoying to work with and required alot of overhead.

What I did

  • Built reward systems for locomotion training.
  • Worked in a Python, PyTorch, and ROS-based training pipeline using PPO.
  • Helped move the simulation workflow from Isaac Sim to Genesis.
  • Worked toward getting behavior onto the real robot, not just keeping it in simulation.
  • Got the robot to walk using the trained model.

What came out of it

  • Cut training time down to roughly five to ten minutes after the Genesis migration.
  • Helped support a much faster sim-to-real workflow for the team.
  • Contributed to a real robot demo.