CoRL 2026 · Accepted
FlowVLA
Enabling World Dynamics Understanding
through Pluggable 3D Flow Learning
1 Overview
2 Abstract
Recent advances in 3D reconstruction enable the extraction of temporally consistent 3D motion from RGB observations; however, current Vision-Language-Action (VLA) models still predominantly rely on 2D image-text priors. Therefore, they do not accurately represent the fundamental dynamics of the 3D world associated with physical actions. This over-reliance on static 2D visual cues limits the capacity of VLA models in achieving robust robot manipulation. To address these limitations, we introduce FlowVLA, a reconstruction-guided, pluggable 3D flow learning framework aimed at empowering VLA models with an inherent understanding of action-relevant dynamics. Instead of requiring explicit 3D sensor inputs or altering the inference-time architecture, our method is designed to utilize temporally aligned 3D flow only during training. Our method allows the model to learn the dynamic evolution of the physical scene while preserving the computational efficiency of conventional 2D VLA backbones. To bridge scene-level dynamics with robotic manipulation, we propose a modular 3D flow learning framework defined by a dual-level flow supervision design. Specifically, point-level flow guides the model to forecast the future displacements of task-critical 3D points, while TCP-level flow aggregates these motions to align with the robot’s Tool-Center-Point (TCP) trajectories. This dual-level flow supervision encourages the model to focus on semantically meaningful dynamic regions, effectively coupling scene-level motion with precise control signals. Extensive evaluations across simulation and real-world experiments demonstrate that FlowVLA consistently improves manipulation performance, zero-shot robustness, and optimizes data efficiency. FlowVLA adds only lightweight auxiliary heads during training and discards them at inference. Therefore, FlowVLA introduces no additional 3D sensor inputs, architectural modifications, or inference-time computation.
3 Method
Paper · pp. 3–5 ↗Reconstruct 3D trajectories from RGB demonstrations, align them to the robot, and select moving points. A point-level head predicts their displacement; a TCP-level head learns to aggregate these motions into the displacement of the robot’s tool-center point.
Scroll to explore ↔View full figure
4 Real-World Demos
Paper · p. 8 ↗Six demonstrations of Seer + FlowVLA on a UR5e.
5 Results
5-a Simulation
Average (4 suites): Seer 87.3% → 92.4%; π₀.₅ 96.8% → 97.5%.
| Method | Spatial | Object | Goal | Long | Average |
|---|---|---|---|---|---|
| Diffusion PolicyRSS 2023 | 78.3 | 92.5 | 68.3 | 50.5 | 72.4 |
| OpenVLACoRL 2024 | 84.7 | 88.4 | 79.2 | 53.7 | 76.5 |
| π₀RSS 2025 | 96.8 | 98.8 | 95.8 | 85.2 | 94.2 |
| CoT-VLACVPR 2025 | 87.5 | 91.6 | 87.6 | 69.0 | 83.9 |
| π₀-FASTRSS 2025 | 96.4 | 96.8 | 88.6 | 60.2 | 85.5 |
| SpatialVLARSS 2025 | 88.2 | 89.9 | 78.6 | 55.5 | 78.1 |
| OctoRSS 2024 | 78.9 | 85.7 | 84.6 | 51.1 | 75.1 |
| UniVLARSS 2025 | 96.5 | 96.8 | 95.6 | 92.0 | 95.2 |
| SeerICLR 2025 | 84.2 | 94.8 | 89.8 | 80.4 | 87.3 |
| Seer + FlowVLA | 93.5 | 96.5 | 91.8 | 87.8 | 92.4 |
| π₀.₅CoRL 2025 | 98.8 | 98.2 | 98.0 | 92.4 | 96.8 |
| π₀.₅ + FlowVLA | 99.0 | 98.8 | 97.4 | 94.8 | 97.5 |
5-b Real World
Average (7 tasks): Seer 61.9% → 79.0%.
Scroll for all tasks
| Base tasks | 3D-aware tasks | Variations | |||||||
|---|---|---|---|---|---|---|---|---|---|
| Model | Cup & Plush | Fruit | Stack Cups | Cabinet | Basketball | Hang Cup | Ring Insertion | Basketball viewpoint | Ring height |
| Seer | 60.0 | 66.7 | 73.3 | 60.0 | 60.0 | 46.7 | 66.7 | 0.0 | 46.7 |
| Seer + FlowVLA | 73.3 | 73.3 | 93.3 | 73.3 | 80.0 | 86.7 | 73.3 | 46.7 | 73.3 |
6 Citation
@inproceedings{go2026flowvla,
title = {FlowVLA: Enabling World Dynamics Understanding
through Pluggable 3D Flow Learning},
author = {Junhyeong Go and Mingyu Jung and Woobin Im and Jongbin Ryu},
booktitle = {Conference on Robot Learning},
year = {2026},
url = {https://gnaroshi.dev/mmai-lab/flowvla/}
}