← Key papers

Paper review · November 15, 2026

A video-action model for data-efficient industrial manipulation

Video-Action Models for Data-Efficient Industrial Manipulation

Robotics & physical AIMultimodalManufacturingMobility Advanced cs.RO

FLUX-mimic attaches an Action Expert to a backbone pretrained on large-scale video and trains on video and actions jointly. It performed automotive manufacturing manipulation tasks with only a small amount of demonstration data.

Problem

Automotive manufacturing tasks such as part insertion, cable routing and soft-part kitting require precise reasoning about contact and geometry. However, existing general-purpose robot datasets focus on household objects and handle such tasks poorly. In addition, specialized equipment and parts make it hard to collect many demonstrations with action labels.

Method

The multimodal foundation model FLUX 3 was used as the video backbone. This backbone was pretrained on web video and in-house industrial data (31 tasks, 295.8 hours), which includes teleoperation video and wearable-camera video without action labels. An Action Expert with the same depth as the backbone references the video features at each layer to generate robot action chunks. Policy training uses a joint video-action approach that applies a flow matching loss to both actions and video. At inference, only one video denoising step is run to obtain features; the full video is not generated.

Results

Limitations

Evaluation was done in a dedicated cell with uniform lighting and background. The scope is also limited to four automotive manufacturing tasks and two platforms. The sample-efficiency experiment and the video-loss ablation were run only on the soft-part kitting task. Because the benchmark task videos were also included in backbone pretraining, generalization to unseen tasks was not separately tested. The provided text was cut off partway, so only some of the detailed numbers could be checked.

Why it matters

On manufacturing floors, it is hard to collect a lot of demonstration data for each task. Using real automotive parts and processes, this study showed that video pretraining can greatly reduce the amount of action data needed. It can serve as a reference when planning data collection for applying general-purpose robot policies to factory automation.

Key terms

VAM(Video-Action Model)
A policy model that predicts robot actions based on the representations a generative video model builds while predicting how a scene changes.
VLA(Vision-Language-Action)
A model that adds robot action outputs on top of a vision-language pretrained model. Examples include π0.5 and OpenVLA.
Flow Matching
A generative method that learns the velocity field of a path from noise to data. In this paper, it was used to generate both video latents and action chunks.
Action Expert
A separate module that takes backbone features and robot state as input and generates a continuous action sequence (action chunk).

Original

Authors: Eugenio Chisari, Stefanos Charalambous, Oliver Sanchez, Thomas Umbach, Michael Gentner, Fabio Quattrini, Norica Bacuieti, Sabin Grube Doiz, et al. · Posted 2026-11-11 · License arXiv 비독점 배포 라이선스 · Based on the full text

This review is an AI-written summary of the paper. Please treat the original paper as the reference.