AICloser-to-target action priors reportedly show no benefit in fine-tuning pretrained robot policies
A research team reports a negative result across 100,000-plus simulation rollouts and 1,250 hardware rollouts, testing a change that prior work had found helpful for policies trained from scratch.
TLDR
The team reports that closer-to-target, non-Gaussian action priors did not improve fine-tuning of Large Behavior Models—robot policies pretrained on large, diverse datasets. It says prior work found that replacing Gaussian priors with closer-to-target alternatives substantially improved diffusion and flow-matching policies trained from scratch. The team hypothesized that the benefit would carry over to fine-tuning, but reports that it did not across 100,000-plus simulation rollouts and 1,250 hardware rollouts.
