FlowCorrect: Efficient Interactive Correction of Generative Flow Policies for Robotic Manipulation

Edgar Welte; Yitian Shi; Rosa Wolf; Maximillian Gilles; Rania Rayyes

cs.ROFeb 25, 2026

FlowCorrect: Efficient Interactive Correction of Generative Flow Policies for Robotic Manipulation

Authors: Edgar Welte, Yitian Shi, Rosa Wolf, Maximillian Gilles, Rania Rayyes

Organizations: Institute for Material Handling and Logistics (IFL), Karlsruhe Institute of Technology, 76131, Germany.

Abstract

Generative manipulation policies can fail catastrophically under deployment-time distribution shift, yet many failures are near-misses: the robot reaches almost-correct poses and would succeed with a small corrective motion. We propose FlowCorrect, a modular interactive imitation learning approach that enables deployment-time adaptation of flow-matching manipulation policies from sparse, relative human corrections without retraining. During execution, a human provides brief corrective pose nudges via a lightweight VR interface. FlowCorrect uses these sparse corrections to locally adapt the policy, improving actions without retraining the backbone while preserving the model performance on previously learned scenarios. We evaluate on a real-world robot across four tabletop tasks: pick-and-place, pouring, cup uprighting, and insertion. With a low correction budget, FlowCorrect achieves an 80% success rate on previously failed cases while preserving performance on previously solved scenarios. The results clearly demonstrate that FlowCorrect learns from very few demonstrations and enables fast, sample-efficient, incremental, human-in-the-loop corrections of generative visuomotor policies at deployment time in real-world robotics.

Figures & tables

Fig. 1: Overview of our interactive FlowCorrect framework: A flow matching visuomotor policy is trained from offline demonstrations. During deployment, the frozen base policy runs the robot while a human provides occasional relative corrections. These sparse corrections are used to train our proposed lightweight FlowCorrect module that locally steers the policy’s flow field, yielding an adapted policy without retraining the original backbone.

Fig. 2: (a) Overview of FlowCorrect module that is attached to the DiTX-Transformer from ManiFlow [ 5 ] : we extend an existing flow matching policy $\pi_{\theta}$ based on DiTX-Transformer with our FlowCorrect module. Our lightweight FlowCorrect module consists of LoRA adapters (parametrized by $\Delta\theta$ ) injected into the transformer, and a gating module $g_{\psi}$ that outputs a signal to steer the vector flow field towards the corrected action. (b) Intuition: across N=4 integration steps, FlowCorrect iteratively adjusts the predicted velocities from $v_{n,t}$ to $v_{n,t}^{*}$ , steering the rollout from a base action $\hat{\boldsymbol{a}}_{t}^{\text{base}}$ toward a corrected action $\boldsymbol{a}_{t}^{\text{corr}}$ .

Fig. 3: Pipeline of the interactive correction interface.

Fig. 4: Hardware setup and representative real-world tasks used in the experiments.

Fig. 5: Top row: Selected ID-hard and OOD-hard initial conditions for the four tasks (left to right): Pouring, Cup Uprighting, Pick-and-Place, and Insertion. The green regions indicate the workspace areas covered by the demonstrations. Middle row: Representative failure cases of the base policy under these conditions. Bottom row: Qualitative examples of successful executions after FlowCorrect fine-tuning on conditions that previously failed.

Panel of Figure 6 — Fig. 6: Overall success rate across 30 ID positions in the workspace. Gray denotes the base policy; red denotes our FlowCorrect (FC) policy after correcting only three ID conditions; and green denotes the base policy retrained (RT) with the same three ID corrections. Notably, FlowCorrect improves the original performance without regression.

Task	Policy	ID- hard1	ID- hard2	ID- hard3	OOD- hard	$\Sigma$
Pick-and- Place	Base	0/10	0/10	0/10	0/10	0/40
	FC (Ours)	3/10	10 /10	9/10	10 /10	32/40
	RT	10 /10	10 /10	10 /10	10 /10	40/40
Pouring	Base	4/10	0/10	0/10	0/10	4/40
	FC (Ours)	10 /10	10 /10	10 /10	2/10	32/40
	RT	10 /10	10 /10	10 /10	10 /10	40/40

TABLE I: Stress-test success on selected hard positions. “ID-hard” are the three low-performing ID conditions; “OOD-hard” is the selected OOD condition.

Mode	Avg. GPU Memory Usage (GB)	Avg. Runtime (min.)
Base	18.84 $\pm$ 0.24	80.86 $\pm$ 10.01
FC (Ours)	4.35 $\pm$ 0.15	30.24 $\pm$ 5.45
RT	19.23 $\pm$ 0.25	52.93 $\pm$ 10.96

TABLE II: Resource usage comparison between our FlowCorrect training ( FC ) and retraining ( RT ).

Variant	ID-30 Avg	ID-hard Avg	OOD-hard Avg
Base	49.76	3.33	0.00
FC (full)	65.00	85.83	77.50
FC w/o gate	54.17	69.17	75.00
FC w/o rollouts	55.84	80.00	87.50
RT	57.50	90.83	95.00
RT w/o rollouts	36.67	90.83	87.50

TABLE III: Ablation study on FlowCorrect . Report success (in %) on 30 ID positions across all tasks, after applying corrections to four hard cases.

Robot Policy Adaptation

Robot policy adaptation changes or steers learned robot behaviors to handle new tasks, environments, hardware, and user preferences.

See all 173 papers

FlowDAgger: Human-in-the-Loop Adaptation of Generative Robot Policies in Latent Space
Jul 9, 2026Michael Murray, Daphne Chen, Simran Bagaria +7Latent Action Learning Robot Policy Learning
1Microsoft Research · University of Washington · 2Microsoft +1
RedFlow: Redirect Failure into Action-level Corrections for Flow-matching VLA Policy
Jul 30, 2026Zhengyang Yan, Junhao Li, Fangqi Zhu +6Robot Policy Learning Policy Optimization
The Hong Kong University of Science and Technology
SeFA-Policy: Fast and Accurate Visuomotor Policy Learning with Selective Flow Alignment
Nov 11, 2025Rong Xue, Jiageng Mao, Mingtong Zhang +1Rectified Flow Visuomotor Policy Learning
University of Southern California

Jul 9, 2026J/K move · Enter open · S save

FlowDAgger: Human-in-the-Loop Adaptation of Generative Robot Policies in Latent Space

Michael Murray, Daphne Chen, Simran Bagaria, Dean Fortier+6

1Microsoft Research · University of Washington · 2Microsoft+1

Latent Action Learning Robot Policy Learning Robot Policy Adaptation Robot Skill Learning Robot Foundation Models Data-Efficient Robot Learning+2

Open paper AskPDF arXiv

Pretrained generative robot policies based on flow matching and diffusion have achieved impressive results across a wide range of manipulation tasks. Yet real-world deployments routinely expose failure modes outside the pretraining distribution. Closing these gaps typically requires large-scale data collection or online reinforcement learning on physical hardware, which is impractical for rapid and safe adaptation. We present FlowDAgger, a sample- and compute-efficient method for adapting frozen generative robot policies from human interventions in latent space. Our key idea is action inversion: each human expert action is mapped to the noise that would have produced it under the frozen base policy, using reverse-time integration followed by local refinement. The resulting inverted noise provides supervision for a lightweight latent policy that steers the base model at deployment time, enabling rapid skill acquisition while preserving its behavioral priors. We evaluate FlowDAgger in simulation and on real-world bimanual and single-arm manipulation, adapting both action-head VLAs and world-action models from a handful of interventions. FlowDAgger outperforms supervised fine-tuning and latent-space RL baselines and preserves pretrained skills on held-out tasks, offering a practical path for adapting robot foundation models in the real world. Website: https://microsoft.github.io/FlowDAgger

Abstract

Robot Policy Adaptation

Explore similar work

FlowDAgger: Human-in-the-Loop Adaptation of Generative Robot Policies in Latent Space

RedFlow: Redirect Failure into Action-level Corrections for Flow-matching VLA Policy

SeFA-Policy: Fast and Accurate Visuomotor Policy Learning with Selective Flow Alignment