Hello everyone,
I have published a self-paced course that connects ROS 2 with today’s AI methods. Everything runs on one simulated greenhouse robot, with one task: inspect the tomato rows.
What you build
- ROS 2 basics: nodes, topics, services and actions
- ROS 2 skill servers that an AI agent (LLM) calls as tools
- A dataset of 49 human demonstrations (4,228 frames)
- A SmolVLA model trained on that data, used as the robot’s “pick” skill
- An honest test of the model, with simple baselines to compare against
What you need: Ubuntu 24.04 and Python. Modules 0 to 2 run on 8 GB of RAM with no GPU. Training the VLA model needs an NVIDIA GPU with 4 GB of memory.
Code: the full workspace (robot, skill servers, agent scripts, demos) is on GitHub and linked in module 0.
Course: Agentic Robotics: From LLM Tool-Calling to Vision-Language-Action Models
Feedback on the design is very welcome.