Four robot policies were trained on 12 to 200 demonstrations of one cheap pick-and-place task, and the authors also measured how noisy a ten-episode block really is
AspireAI community · arXiv
ACT, Diffusion Policy, SmolVLA and pi-0.5 were trained on nested subsets of 12, 25, 50, 100 and 200 teleoperated demonstrations of a single pick-and-place task with distractor objects, one fixed training recipe per policy, then run on a physical SO-101 arm over 1,340 scored episodes in 126 blocks. ACT trained from scratch went from 1 success in 40 episodes at 12 demonstrations to 37 in 40 at 100, while pi-0.5 with its action expert fine-tuned led when data was scarce at 16 of 40 at 12 demonstrations and 28 of 40 at 25, stayed near 70 percent, and from 100 demonstrations ACT was ahead by a margin the study does not fully resolve. SmolVLA and Diffusion Policy peaked near 50 percent and both were hit by a failure to start, with SmolVLA's arm never moving in 47 percent of episodes; for pi-0.5 the fine-tuning recipe mattered more than data volume, as low-rank adaptation succeeded in 4 of 110 episodes against 130 of 200 for action-expert fine-tuning. The measurement contribution matters as much as the results: the same checkpoint tested twice differs by 2 successes in 10 on average, so a ten-episode block is a weak signal, and the paper is explicit about which conclusions survive the confound that data amount and number of epochs moved together.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.