Video-Action Models for Data-Efficient Industrial Manipulation

Eugenio Chisari, Stefanos Charalambous, Oliver Sanchez, Thomas Umbach, Michael Gentner, Fabio Quattrini, Norica Bacuieti, Sabin Grube Doiz, Valentin Bertle, Sigmund Hennum Hoeg, Alain Ryser, German Rodriguez, Mario Arduini, Dang Nguyen, Florian Aumann, Felix Hegg, Manuel Meier, Philipp Wand, Julian Lotzer, Alessio Desogus, Benedek Forrai, Michael Bloesch, Stephan-Daniel Gravert, Stefan Weirich, Patrick Lanners, Elvis Nava

November 2026

Control unit assembly on the bimanual Franka with dexterous hands, autonomous at 1x speed.

TL;DR FLUX-mimic attaches an action expert to a large pretrained video model and trains it on video and actions from automotive assembly tasks. On a bimanual Franka with dexterous hands it averages 82.8% success over four factory tasks, against 52.6% for a single-task policy and 27.6% for pi0.5.

What they do

Industrial tasks such as inserting parts into tight fixtures, routing wire harnesses and kitting foam parts look nothing like the household data that generalist robot policies are trained on, and demonstrations are expensive to collect. FLUX-mimic, from Mimic Robotics and the BMW Group, tests whether a pretrained video model can make up for scarce action data. They collected 295.8 h of industrial data over 31 tasks (218.9 h robot teleoperation, 76.8 h from camera-only wearable gloves) and evaluated on four automotive tasks on two bimanual robots.

How it works

Video backbone. Initialized from FLUX 3, pretrained on millions of hours of general video. The whole industrial dataset, benchmark tasks included, is added to its pretraining.
Action expert. As deep as the backbone, with each layer attending to the matching backbone layer, plus the robot state. At inference one video denoising step supplies the features; no full video is generated.
Joint training. Video and action are both trained with flow matching at equal weight. Actions are predicted at 15 Hz in chunks of 20 and executed with real-time chunking.

Results

Franka arms with dexterous hands: 85% assembly, 85% fixture insertion, 100% soft part kitting, 61.1% wire harness routing, an average of 82.8% against 52.6% for a single-task flow-matching policy and 27.6% for pi0.5.
Trossen arms with grippers: both FLUX-mimic and pi0.5 average 60% over two tasks.
Soft part kitting with 1 h of demonstrations: 45% versus 7.5% and 2.5% for the baselines. With 8 h it reaches 95%, above both baselines trained on the full 32 h (70% and 82.5%).
Dropping the video loss during policy training lowers kitting success from 20/20 to 13/20.

Limitations

The paper has no limitations section. Evaluation runs in a dedicated cell with uniform lighting and background, with 18-20 rollouts per task. The benchmark recordings are also part of the backbone’s pretraining, and the data-budget and video-loss studies cover only soft part kitting. On the gripper robot it does not beat pi0.5.

Videos

Soft part kitting, picking foam strut covers from a bin into ordered slots.

Wire harness routing, clipping a cable into the clips on a door panel.

Figures

From Figure 3: FLUX-mimic architecture. The video backbone is conditioned on the task caption and past video frames, ... — From Figure 3: FLUX-mimic architecture. The video backbone is conditioned on the task caption and past video frames, the action expert on the robot state. After one video denoising step, the action expert attends layer by layer to the backbone features to predict robot actions; training uses a joint video-action loss.

From Figure 2: Recorded time per task in the industrial data collection, for the four benchmark tasks (left) and 27 a... — From Figure 2: Recorded time per task in the industrial data collection, for the four benchmark tasks (left) and 27 additional pretraining tasks (right). Bars are stacked by source: wearable recording, Franka teleoperation and Trossen teleoperation.

From Figure 5: Sample efficiency on soft part kitting with Franka. Success rate against action-data budget from 1 h t... — From Figure 5: Sample efficiency on soft part kitting with Franka. Success rate against action-data budget from 1 h to 32 h, each checkpoint evaluated over 40 rollouts; shaded lobes show the posterior over the true success rate.

Abstract

Industrial manipulation requires precise alignment and controlled handling of rigid, deformable, and flexible components. Yet, the benefits of pretrained visuomotor policies remain underexplored in settings where limited access to specialized hardware, components, and fixtures constrains demonstration collection and evaluation. Understanding task-specific data requirements is important for assessing their scalability across diverse processes on the factory floor. We investigate whether video pretraining can support reliable industrial manipulation with limited action data. We propose FLUX-mimic, a video-action model, motivated by the potential of video pretraining to capture motion and interaction patterns that enable sample-efficient training. FLUX-mimic is built on a backbone pretrained on large-scale web and industrial task data, and an action expert to generate robot trajectories. We train our policy by joint video-action training. Across four automotive manufacturing benchmark tasks and two bimanual robot platforms, FLUX-mimic outperforms all considered baselines on the dexterous robot platform and matches pi_{0.5}'s average success rate on the gripper-based robot platform. On a deformable part kitting task, FLUX-mimic reaches 77.5% success with as little as two hours of task-specific action-labeled demonstrations.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.

Matt Strong