Courses / programming

Deep RL for Robotics

advanced~16 hours12 lessonsEnvironment: docker

Prerequisites: reinforcement-learning, ros2-basics, gazebo-sim, ros2-navigation

Start: Gymnasium: the environment interface everyone agrees onPractice questions

Modules

Module 1 — The modern stack
  1. 1Gymnasium: the environment interface everyone agrees on
  2. 2stable-baselines3: PPO/SAC without writing them again
  3. 3Wrapping YOUR robot as a Gym environment
Module 2 — Training that finishes
  1. 1Why we do NOT train in Gazebo (and what we train in instead)
  2. 2A fast headless sim for a diff-drive robot
  3. 3Training run: PPO on a navigation task
  4. 4Reading the curves: reward, entropy, value loss, and when to kill a run
Module 3 — From policy to robot
  1. 1Evaluating the policy in Gazebo
  2. 2Wrapping the policy as a ROS 2 node (obs from /scan + /odom, out to /cmd_vel)
  3. 3The safety envelope: never let a neural network drive unsupervised
Module 4 — The verdict
  1. 1Learned vs classical: benchmark your PPO policy against Nav2
  2. 2Capstone: a learned local planner with a classical safety layer