FlowDAgger
Official implementation of FlowDAgger: Human-in-the-Loop Adaptation of Generative Robot Policies in Latent Space, published at the 10th Conference on Robot Learning (CoRL 2026).
Project page: https://microsoft.github.io/FlowDAgger
FlowDAgger is latent-space DAgger for flow-matching and diffusion based robot policies. Instead of fine-tuning the base policy, it learns a small steering network that outputs an observation-conditioned initial latent for the policy's sampler. Expert corrections are mapped back into the sampler's initial-latent space by inverting the policy's sampling ODE, and the steering network is trained to match those inverted targets with a behavior-cloning loss.
This repo contains minimal reference implementations for two base-policy backends. Each backend has its own dependencies and runnable example.
How this example works
- Roll out the base policy. A steering network outputs an observation-conditioned initial latent for the sampler.
- An intervention handler hands control to a scripted expert.
- Each expert action chunk is inverted through the policy's sampler to estimate the initial latent that would have produced it.
- The steering network is trained to match those inverted latent targets (MSE).
- Repeat. Over time the steering network reproduces expert behavior without touching the base-policy weights.
Layout
shared/ scripted expert, task registry, intervention handler
flowdagger_pi05/ JAX, pi0.5 base, MetaWorld assembly
(openpi is a git submodule under flowdagger_pi05/openpi)
flowdagger_gr00t/ PyTorch, GR00T N1.7 base, LIBERO-90 task 57
Getting started
The backends use separate Python environments because their JAX and PyTorch stacks have different dependencies. Follow the README for the backend you want to run:
The pi0.5 backend uses openpi as a git submodule:
git submodule update --init flowdagger_pi05/openpi
Citation
@inproceedings{murray2026flowdagger,
title={{FlowDAgger}: Human-in-the-Loop Adaptation of Generative Robot Policies in Latent Space},
author={Murray, Michael and Chen, Daphne and Bagaria, Simran and Fortier, Dean and Hellebrekers, Tess and Mullins, Galen and Gajarla, Harshavardhan and Mees, Oier and Cakmak, Maya and Kolobov, Andrey},
booktitle={Proceedings of the 10th Conference on Robot Learning},
year={2026}
}License
MIT. See LICENSE.