Guides · Updated 2026-08-23
Everyone asks for a magic episode count. There is not one, and anyone who gives you a single number is guessing. What exists are honest reference points from real projects.
SmolVLA, the community scale vision language action model from the LeRobot team, pretrained on 481 community contributed datasets holding 22.9K episodes and 10.6M frames in total, then gets fine-tuned per task. At the hobby end, people routinely fine-tune a single pick and place behavior with tens of carefully recorded episodes on one arm. Between those two points is where most projects live.
Three things, in order. First, labels: language conditioned training needs real task sentences, not placeholders. Second, consistency: same camera names, stable framing, complete metadata. Third, coverage: vary object positions and lighting across episodes instead of recording the same motion fifty times.
Record a first batch of clean episodes for your task. Run datum-lint so labels and cameras are right before training, not after. Train, watch failure cases, then record more episodes that target exactly what failed. Iterating beats stockpiling.
The Datum finder filters 1000 checked public datasets by robot, size, labels, and license, with junk hidden. Co-training on clean public data plus your own episodes is standard practice now.