AMD AI Dev Day robotic arm RL workshop

This is the notebook "Build a Synthetic Data Generation Pipeline in a Digital Twin World". We train a Franka Panda robotic arm with a SmolVLA model. The synthetic kitchen scene is generated with https://marble.worldlabs.ai/ and pregenerated and provided by the instructor (along with all the training scripts, the model, the iPython notebook in the Jupyter Lab environment. I'm mentioning Gemma 4, actually Google has X-VLA model but that requires some adjustments. The first training yields a model which constantly positions the beak a little to the right (when viewing form up) and misses all 10 of the simulations. Then I retrain with 10000 steps (up from 4000) and that model has 3 successful pickups, and one where it picks up the red cube but it falls down (I'd consider that half success). During the conference I was lucky and my 4000 step trained model had 6 success out of 8 episodes.