Diffusion Policy for Interactive Fleet Learning with Scalable Human Supervision 37 views

In large-scale robot fleet deployment, a small number of human supervisors must oversee many robots acting in parallel. Interactive fleet learning (IFL) addresses this setting in which multiple robots can query limited human supervisors while intervention data are aggregated over time to update a policy. A central challenge in this setting is how to allocate limited supervision and determine which queried states should be selected for intervention and learning. Existing implementations of the IFL benchmark instantiate the imitation learner as a multi-layer perceptron (MLP), which is limited in modeling multimodal action distributions. Motivated by prior work that uses diffusion loss as a query signal in single-robot imitation learning, we use diffusion policy (DP) and deploy diffusion loss as an uncertainty signal for threshold calibration, intervention prioritization, human allocation, and continual policy improvement in online fleet learning. We evaluate the resulting system on Allegro Hand. The IFL with diffusion policy achieves higher episode success rates than the MLP baseline, and our threshold diagnostic shows that outcome-based recalibration makes the diffusion-loss query rule more budget-aware under limited supervision.

MS (Master of Science)

English

All rights reserved by the author (no additional license for public reuse)

2026-04-21