Keywords: Robot-gated DAgger, Diffusion Policy
TL;DR: Diff-DAgger enhances Diffusion Policy through data aggregation, effectively handling multi-modality in demonstration data. Unlike Ensemble DAgger, it provides a more reliable approach for a robot's self-assessment of uncertainty.
Abstract: Abstract: Diffusion policy has shown impressive results in handling multi-modal
tasks in robotic manipulation. However, it has fundamental limitations in out-of-
distribution failures that persist due to compounding errors and its limited capa-
bility to extrapolate. One way to address these limitations is robot-gated DAgger,
an interactive imitation learning with a robot query system to actively seek ex-
pert help during policy rollout. While robot-gated DAgger has high potential for
learning at scale, existing methods like Ensemble-DAgger struggle with highly
expressive policies: They often misinterpret policy disagreements as uncertainty
at multi-modal decision points. To address this, we introduce Diff-DAgger, an ef-
ficient robot-gated DAgger algorithm that leverages the training objective of dif-
fusion policy. We evaluate Diff-DAgger across different robot tasks and show that
Diff-DAgger overall achieves 81.5% accuracy on the task failure prediction and
improves task completion rate by 13.8%. We hope that this work opens up a path
for efficiently incorporating expressive policies into interactive robot learning.
Submission Number: 43
Loading