FlowDAgger: Robots Learn via Corrections
oier_mees · x · 2026-07-16
Core Conclusion
The authors introduce FlowDAgger: instead of directly fine-tuning a robot foundation model, it leverages human corrective actions to learn how to "steer" an already frozen policy.
Key Methods
- Modern VLA / world-action models struggle to adapt to new robots and tasks.
- Traditional approaches like online imitation learning or retraining often break existing capabilities when data is scarce.
- FlowDAgger uses action inversion to map human corrective actions back into the latent noise space of the frozen generative policy.
- A lightweight controller is then trained using these latent targets for deployment-time adaptation, leaving the base model untouched.
Results
- Requires only 5–20 human interventions to learn.
- Outperforms supervised fine-tuning and latent-space RL in both simulation and real-world robots.
- Applicable to VLA, diffusion policies, world-action models.
- Delivers stable improvements without modifying the pre-trained policy.
Additional Info
This project is a collaboration between Microsoft Research and the University of Washington, with paper, project page, and code available.
Related event: FlowDAgger: Adapting Robot Models via Human Corrections(2 posts)→
More from Embodied
- Purdue's autonomous drone tracks runners through dense forests with on-chip AI, no GPS — birchlse · 2026-10-27
- Skyfall AI's SKY-001 world model turns sketches into CAD builds in 3.9s, vs 199s for GPT-6 Astra — rohanpaul_ai · 2026-10-27
- Tesla FSD approved in Latvia, the 10th European country; Musk praises smaller nations for moving fastest — elonmusk · 2026-10-27
- WanGP v18.81 adds Tencent's Hy-Nomas world model, playable on 8GB VRAM — cocktailpeanut · 2026-10-27
- NVIDIA open-sources FLORA, a robotics data engine ~125x cheaper than GPT-6 Astra — nvidia · 2026-10-27
- Embodied AI startups need a different operational discipline, investor says — ___Mufasaa · 2026-10-27