| Project Domains | Mentors |
|---|---|
| Reinforcement Learning, Deep Learning, Bipedal Locomotion, Robotic Simulation | Gargi Gupta, Abhijeet Bhalerao |
Project Description
Robots don’t intuitively walk; they’re taught using a policy. Waddle teaches you how to build that policy. You’ll learn to implement reinforcement learning algorithms by solving mini environments. Then you will apply them in MuJoCo and Isaac Lab to train Open Duck, an open-source bipedal robot. The core of the project is reward engineering i.e. taking a working locomotion reward apart term by term, then writing your own and watching it learn to balance and walk. The final stretch benchmarks PPO against SAC on the same robot, adds velocity-command control, and stress-tests the policy with pushes and rough terrain.
Resources
What is Reinforcement Learning?
Introduction to Proximal Policy Optimization
Universal RL Textbook
RL for locomotion