Both simulation and real robot experiments use expert demonstrations collected using scripted policies, except for Charger Plug, where demonstrations are generated by an externally trained reinforcement learning policy. To obtain imperfect but reproducible agents, we use two types of base policies.
Experimental Setup
Both simulation and real robot, expert demonstrations are collected using scripted policies, except for Charger Plug, where demonstrations are generated by an externally trained RL policy. To obtain imperfect but reproducible agents, we use two types of base policies.
Result Analysis
Overall, [a flow-based action adaptation method] provides robust improvements across agents and environments. Among the 16 task-wrapper settings, our variants achieve the best success rate in 13 cases and remain competitive in the others, showing that action-level adaptation is effective across delayed, noisy, slow, and biased agent behaviors.
Specifically, FPAS alone already provides substantial gains over the source agents in most settings and remains competitive with strong baselines, highlighting the effectiveness of the flow-based OOD detection itself: even without additional generative correction, it provides a useful signal for identifying and refining problematic actions. Compared with FPAS, FEEG provides additional gains in several challenging settings, especially on high-dimensional manipulation tasks such as Can and Charger, with the largest improvement exceeding [a large performance improvement]. These results suggest that flow-guided action editing can produce more effective corrections than local sampling when successful adaptation requires coordinated changes across the action chunk.
IFAE shows the clearest advantage when the agent actions form a source distribution that can be modeled reliably. In the stochastic-wrapper experiments, IFAE achieves the best or near-best performance in several settings. These gains suggest that source-to-expert flow editing is effective when the wrapped agent still preserves a coherent action manifold and the required correction can be expressed as a distribution-level transport. However, IFAE is less reliable under some settings that the agent's action is less likely captured by the learned source flow. In these cases, FEEG is often stronger because it relies primarily on the expert flow and a proximity term, rather than requiring an accurate model of the corrupted source distribution.
Additionally, the OOD detection provides an efficient mechanism for selective intervention. Instead of refining every proposed chunk, [flow-based methods] apply correction only when the OOD score indicates that the chunk is outside the expert distribution. This reduces inference cost and avoids perturbing already valid actions while using the same expert flow for both OOD detection and correction.
Conclusion
We presented [a family of flow-based action-adaptation methods] that improves imperfect agent actions without updating the underlying agent. Using limited expert demonstrations, [this method] supports both policy adaptation and shared autonomy, and uses the learned expert flow for selective OOD-gated intervention. Experiments across simulation and real robots show consistent improvements for stochastic-wrapper imitation learning and VLA agents.
Limitations
Our experiments show that FEEG and IFAE provide complementary adaptation behaviors, but the factors that determine which variant is best for a given task remain to be studied more systematically. Future work should further analyze the effect of editing hyperparameters and task properties such as contact richness, precision requirements, and compositional structure. More extensive ablations and automatic variant selection may further improve the robustness of flow-based action adaptation.
Frequently Asked Questions
What is the main innovation of flow-based policy adaptation? It improves imperfect agent actions at inference time without retraining or fine-tuning the underlying policy, using a learned flow model trained on expert demonstrations.
How does selective OOD-gated intervention work in this framework? The method only applies corrections when the OOD score indicates an agent action chunk is outside the expert distribution, reducing compute cost while preserving already valid actions.
What are the key differences between the FEEG and IFAE variants? FEEG uses only the expert flow and a proximity term for corrections, while IFAE models source-to-expert distribution transport and works best when the agent's action manifold remains coherent.
In which types of tasks does flow-based action adaptation show the largest improvements? The method provides the strongest gains on high-dimensional manipulation tasks like Can and Charger, where coordinated changes across an action chunk are required for successful adaptation.