Flow as Flow:
Modeling Robot Velocity Fields
as
Probability Velocity Fields

Anonymous Author(s)


Under Review



Overview of Flow as Flow framework.

Overview of Flow as Flow. We leverage diverse cross-embodiment videos from multiple robot embodiments and humans for training. At deployment, the model predicts a robot flow (robot velocity field) representing a task-relevant motion conditioned on an initial image and a goal image. The robot then executes object manipulation conditioned on the generated flow to achieve the object poses specified by the goal image.

Abstract

Cross-embodiment data have become central to training robotic foundation models. To leverage such heterogeneous data, we focus on flow-based object manipulation, where robot flows (robot velocity fields) serve as embodiment-agnostic motion representations. Prior work often formulates robot flows by differencing predicted keypoints across frames, which requires a strong visibility assumption and thus yields rough approximations of their underlying velocity fields. To address this limitation, we propose Flow as Flow, a framework that models robot flows as probability flows based on a flow matching formulation. By naturally modeling such velocity fields within this formulation, our method achieves efficient and high-quality robot flow generation. Across standard benchmarks, our method outperforms representative baseline methods on standard metrics, while achieving approximately 33× faster generation. Furthermore, through real-world experiments evaluating 9 methods with 260 trials per method across 13 manipulation tasks, we show that our method achieves a higher average success rate than the baseline methods.

Highlights

🚀 ~33× faster generation than the best baseline method (44 ms vs. 1,430 ms per sample).

📊 Achieves the best ADE scores across four standard benchmarks (Fractal, Bridge V2, DROID-100, Fanuc Manipulation), including zero-shot settings.

🤖 Achieves the highest average success rate across 13 real-world mobile manipulation tasks, outperforming the strongest baseline (Track2Act) by 10 points.

🔌 Architecture-agnostic: Flow as Flow can be directly integrated into other flow-based methods without any structural changes.

Real-World Experiments

Proposed Framework

Flow as Flow

The core novelty of our framework is modeling physical robot velocity fields as probability velocity fields in the generation space of flow matching.

Proposed method visualization

We initialize \(N\) (e.g., \(10\times 10\)) points uniformly on the image and obtain their future positions by integrating the velocity fields predicted by a flow generation model \(\boldsymbol{v}_\theta\).

We construct target velocity fields in a stabilizing feedback form: \[ \boldsymbol{v}(\boldsymbol{\Xi}_h, \boldsymbol{X}, h) = \dot{\boldsymbol{\Xi}}_h - k(\boldsymbol{X} - \boldsymbol{\Xi}_h), \] where \(\boldsymbol{X} \sim \mathcal{N}(\boldsymbol{\Xi}_h,\, \sigma_0^2 e^{-2kh}\boldsymbol{I})\). The stabilization term enhances robustness to out-of-distribution samples.

We train \(\boldsymbol{v}_\theta\) with the conditional flow matching (CFM) loss: \[ \mathcal{L}_{\text{CFM}} = \mathbb{E}_{\boldsymbol{\Xi}_h, h, \boldsymbol{X}} \left[\left\lVert \boldsymbol{v}_\theta\!\left(\boldsymbol{X}, h \mid \mathcal{I}, \mathcal{G}, \boldsymbol{\Xi}_{0:h-1}\right) - \boldsymbol{v}\!\left(\boldsymbol{\Xi}_h, \boldsymbol{X}, h\right) \right\rVert^2\right]. \] At inference, coordinates at step \(h\) are obtained by integrating \(\boldsymbol{v}_\theta\) autoregressively, enabling fast generation with only a single ODE solve per step: \[ \boldsymbol{X}_h = \boldsymbol{X}_0 + \int_0^h \boldsymbol{v}_\theta\!\left(\boldsymbol{X}_\tau, \tau \mid \mathcal{I}, \mathcal{G}, \mathcal{X}_{<\tau}\right) d\tau. \]

Model Architecture

Model architecture of Flow as Flow.

Our framework consists of two main modules: the Flow Generation Module and the Action Generation Module.

Results

Flow Generation

Quantitative Results

Method Flow as Flow In-domain   Zero-shot   Inf. speed ↓
[ms]
Fractal Bridge V2 DROID-100 Fanuc Manip.
ADE ↓FDE ↓LTDR ↑[%] ADE ↓FDE ↓LTDR ↑[%] ADE ↓FDE ↓LTDR ↑[%] ADE ↓FDE ↓LTDR ↑[%]
Language-conditioned
FLIP 66.1787.5235.69 50.7368.4347.72 43.1049.1054.87 28.3150.8372.17 35
FLIP 38.7757.4158.11 48.3466.3149.26 38.5444.4856.25 26.7947.8571.62 17
Im2Flow2Act 37.1447.7460.61 51.4870.9347.97 44.1454.4151.48 38.1564.1859.25 5,580
Im2Flow2Act 33.2146.8364.25 42.9660.9354.00 38.8745.2556.07 26.5148.4871.75 230
GigaWorld-0-Video 74.0095.2332.46 53.1869.5846.44 42.7547.9653.91 37.6058.3561.22 26,976
Goal-conditioned
Track2Act 64.3286.6242.00 47.2964.1351.61 40.7347.4354.29 27.3747.1770.99 1,430
Ours 21.2327.3176.79 27.1134.6669.96 35.8940.5858.81 22.4642.1974.54 44

Quantitative comparison on robot flow generation benchmarks. Bold indicates the best result and underline indicates the second best. Green rows indicate methods using Flow as Flow (ours = dark green, variants = light green).

Qualitative Results

Qualitative results of robot flow generation.

Qualitative results of robot flow generation. Our method generated flows targeting the correct object with the appropriate motion direction, both in in-domain and zero-shot settings.


Real-World Robot Experiments

Real-world experimental setup with the Human Support Robot.

Representative examples of mobile manipulation tasks. 13 diverse tasks: bin picking, bussing table, push bin into shelf, push chair, open/close drawer, put fruit on plate, close box, water plant, take towel, close laptop, stack block, and stack cup.


Quantitative Results

Scroll to see all tasks
[%] Method Flow-
based
Flow
as Flow
Push bin
into shelf
Push
chair
Close
drawer
Close
box
Take
towel
Bin
picking
Put fruit
on plate
Bussing
table
Close
laptop
Water
plant
Open
drawer
Stack
cup
Stack
block
Avg.
Language-conditioned
DP-Lang 757565653535 353535520510 38
FLIP 455550605025 151525510105 28
FLIP 555570655545 60454025251010 43
Im2Flow2Act 707065656045 50404520201520 45
Im2Flow2Act 858085757060 60555525252015 55
Goal-conditioned
DP-Goal 403045502530 515305505 22
Track2Act 858075606555 55553515152010 48
Ours 909080757570 65655030302015 58
Oracle 959090858080 75705540352520 65

Quantitative results of real-world experiments (260 trials per method, 20 per task). Bold = best, underline = second best. Our method achieved 58% average success rate, outperforming Track2Act (48%) by 10 points.


Qualitative Results

Qualitative results of real-world experiments.

Qualitative results of our method in real-world experiments, showing three successful rollouts: (a) bussing table, (b) close laptop, and (c) push chair.


BibTeX

To be announced.