Alright, I have 25 episodes recorded for my block stacking task, so I decided to train my first test policy using the Action Chunking Transformer (ACT) framework to see how well my data collection has been going.
Figuring out how to train on the data I've collected was a lot easier than collecting the data, in my opinion. First, the command for training using the data was a lot easier to figure out compared to the command for recording the robot's movements, due to having less parameters (7 for the train command vs 15 for the record command). Secondly, because I had already set up all the necessary logins when I started recording data, I had all the repositories ready to go and logging in to wandb (which what lerobot uses to visualize training graphs within the command) was easier because I had done the hugging-face cli login before. Finally, once the training command is running, you can leave it to train and do other things while you wait. It did take a while though (~3 hours with 35k steps) and I stopped the training when my loss started to plateau around 0.059.
When I evaluated the policy, it was about what I expected. The robot arm could sometimes get the center block and left block into their respective boxes, with the rightmost block being less consistent. This reflects the layout I've set up on my workspace: the left and center block are the two closest to the robot arm due to where I put the boxes relative to the arm, while the arm has to reach a bit to put the right block into it's bin. Even with only 25 episodes for referencing, the robot arm's policy still did a lot better than I thought relative to the sample size.
Overall, it was exciting to see the robot start to learn from my demonstrations. There’s clearly a lot of room for improvement, but this first policy gives me a solid baseline to build on as I collect more data — especially recovery demonstrations.