MuJoCo + PyTorch

Move from scripted robot control to imitation learning, vision-language-action policies, PPO, and cross-domain reinforcement learning.

Goals

• Control robosuite environments and design Gymnasium rewards
• Collect expert trajectories and train behavior cloning policies
• Fine-tune SmolVLA on a LeRobot dataset
• Train PPO locomotion and study cross-domain policy transfer

Simulation and Imitation Learning

• MT01 — Robosuite Introduction

  Explore observations, actions, and random rollouts in a Panda Lift task.

• MT02 — Controllers and Cameras

  Script a reach-grasp-lift sequence and inspect visual observations.

• MT03 — Gymnasium and Reward

  Wrap a robot task as Gymnasium and design a dense reward.

• MT04 — Behavior Cloning

  Collect expert data and train a policy to imitate it.

• MT05 — SmolVLA Fine-tuning

  Fine-tune a vision-language-action model for manipulation.

Reinforcement Learning

• MT06 — PPO

  Train an end-to-end PPO locomotion policy.

• MT07 — Cross-Domain RL

  Transfer knowledge across morphology and task changes.

Return to the Physical AI overview.