FlowDAgger adapts robot policies with just a few human demonstrations
New method injects human corrections into frozen robot models using 'action inversion'...
FlowDAgger tackles a key bottleneck in deploying robot foundation models: gaps in pretraining distribution that cause failures. Traditional fixes like large-scale data collection or online RL are impractical. The team's innovation is 'action inversion'—mapping each human intervention to the noise that would generate that action under the frozen base policy. This inverted noise becomes supervision for a lightweight latent policy that steers the model at deployment time, enabling fast skill learning while retaining prior behaviors.
Evaluated in simulation and on real bimanual (two-arm) and single-arm robot setups, FlowDAgger required only a handful of human corrections to adapt policies. It consistently beat supervised fine-tuning and latent-space reinforcement learning, both in adaptation speed and in preserving performance on held-out tasks. The approach works with both action-head VLAs and world-action models, offering a practical path for closing the sim-to-real gap in robot learning.
- Action inversion technique maps human actions to latent noise using reverse-time integration and local refinement.
- Requires only a handful of human interventions to adapt frozen generative robot policies.
- Outperforms supervised fine-tuning and latent-space RL on bimanual and single-arm manipulation tasks.
Why It Matters
Enables rapid, sample-efficient, and safe adaptation of robot foundation models for real-world deployment.