Event

Comparison of Data Scale for Four Imitation Learning Strategies on a Low-Cost Robotic Arm

1 reports · 1 sources

Simon Cowling published an experimental study: on a low-cost SO-101 robotic arm, four imitation learning strategies—ACT, Diffusion Policy, SmolVLA, and π0.5—were trained with 12, 25, 50, 100, and 200 teleoperated demonstrations, with each strategy using a fixed training recipe, and then evaluated over 1340 episodes at fixed object placement positions. The study notes that because training duration is fixed within each recipe, data volume and the number of times the data is repeatedly seen change simultaneously; based on this, the author explains which conclusions are unaffected by this confounding factor and which are affected. For π0.5, the fine-tuning recipe is more critical than data volume: low-rank adaptation succeeded 4 times in 110 episodes, whereas action expert fine-tuning succeeded 130 times in 200 episodes.