Brands
Latest top stories
Technology

Can video footage teach factory robots new tricks faster?

19 November 2026

Humanoid robotics at BMW Group Plant Leipizg. The pictured robot was not part of the study (image: BMW)

 

BMW Group and the Swiss start-up mimic robotics are testing whether video footage can reduce the time operators spend teaching factory robots new jobs. A robot whose model had already learned from industrial footage completed the final stage of a parts-handling task in 31 out of 40 trials after two hours of guided demonstrations, offering BMW an early measure of how much hands-on training time it might save.

The machine had to pick foam strut covers from a bin and arrange them in a narrow row, yellow side down. Because the covers could deform as they were lifted, each one had to be placed into its slot in the correct orientation. To record a demonstration, an operator guides the robot through the motions while cameras capture what the machine sees and the system logs its movements. Because the robot, parts and fixtures must all be available for this process, data collection forms a significant part of the cost of introducing a new operation.

Four BMW Group researchers co-authored the paper, which was first posted as a preprint on 11 November 2026 and revised on 17 November. Using automotive production components, the team tested four operations, ranging from fitting a rigid control unit into a fixture to routing a flexible cable along a door panel. Each places different demands on the machine: while a control unit must line up precisely with its fixture, moving one section of a cable can cause the next to shift.

 

How mimic robotics trains machines with video

 

Mimic emerged from research at ETH Zurich’s Soft Robotics Lab and became a spin-off in 2024. The company develops dexterous hands that can be fitted to conventional robot arms, using wearable cameras to record how people handle parts. In the BMW project, those cameras supplied nearly 77 hours of footage, including 11 operations for which the team collected no guided robot demonstrations.

The researchers built their video-action model, FLUX-mimic, on Black Forest Labs’ FLUX 3. During training, the system learned both to predict how a scene would change and to generate the movements a robot could make. Once working, the system uses what it has learned about those visual changes to calculate movements without rendering a new video at every step.

In total, the project gathered nearly 296 hours of industrial recordings across 31 operations, with the earlier video training covering all four benchmark tasks. The two-hour figure refers specifically to the additional demonstrations used to teach the robot the strut-cover operation. How much practice the model would need for a job it had never seen before remains an open question.

 

What another hour of practice bought

 

The researchers ran 40 trials at each of six training budgets, ranging from one to 32 hours of guided demonstrations. On the final stage of the strut-cover task, success rose from 45% after one hour to 77.5% after two, reaching 95% after eight. At the higher budgets, scores levelled off and fluctuated. For BMW, the experiment offers a way to estimate what another round of demonstrations might yield before committing more time on the equipment.

A separate test examined whether asking the model to predict video during training helped. Starting from the same pretrained video model, the version trained to predict video alongside robot actions completed the full strut-cover task in 20 out of 20 trials, whereas the version trained to predict actions alone succeeded in 13 out of 20. Although limited to a single operation, the comparison isolates that part of the training method.

 

Where the gains narrowed

 

Across the four automotive tasks, FLUX-mimic averaged an 82.8% full-task success rate on two Franka arms fitted with dexterous hands, compared with 52.6% for the stronger of the two comparison methods. Its lead was narrowest on cable routing, where it secured the cable in all three clips in 11 of 18 trials – just one more than that baseline.

On a Trossen platform with simpler grippers, FLUX-mimic and Physical Intelligence’s π₀.₅ both averaged 60% across two control-unit tasks. The models were trained separately for each platform and started from different datasets, with industrial tasks absent from π₀.₅’s published pretraining mix. The figures show how the systems performed under the conditions tested.

MTN’s profile of Jeff Bezos-backed Physical Intelligence explored how its newer π₀.₇ model folded clothes with industrial arms after learning the task on smaller machines. The BMW study focuses on a question that arises when planning a factory installation: how many examples must an operator record for a particular robot, part and job?

For BMW, the strut covers offer a candidate for testing whether a shorter training period can translate into reliable work at production speed. The cable points to a different cost, as even after training, each failed attempt could require an operator to step in. Ultimately, the decision to automate a job depends on both demands on an operator’s time: from the hours spent teaching the robot to the interventions it needs once the line is running.

 

 

Further reading on MoveTheNeedle.news:

Physical Intelligence: the Bezos-backed company teaching robots to adapt

Agility joins Nasdaq as Digit humanoid robots expand