Start here · Train a brain · Updated 2026-08-23
Your first policy will grab air sometimes. Everyone's does. The difference between people who get a working robot and people who quit is what they do with the failures.
Run rollouts and sort what you see into three buckets. Misses in one region of the table: your episodes did not cover that region; the official guidance of about 10 episodes per object position exists exactly for this. Hesitates or oscillates: often operator wobble in the demonstrations; your smoothest episodes teach best. Worked yesterday, fails today: something moved, usually a camera or the lighting; the model memorized the scene more than you think.
Do not blindly record 50 more episodes. Record 10 that target the failure bucket: the missed region, the tricky grasp angle, the new lighting. Resume into the same dataset with --resume=true, retrain, roll out again. One targeted batch usually moves success more than doubling the dataset.
Ten rollouts, count successes, write the number down before and after each retrain. Memory lies about progress; a notebook does not. This tiny habit is also exactly how serious labs measure policies, so you are learning the real skill.
Try the bigger pretrained model (SmolVLA fine-tune instead of ACT from scratch), or co-train with clean public data from the same robot: the finder filters 1000 checked datasets by robot, labels, and license, junk already hidden. Someone else's clean episodes plus yours is standard practice.