Robotics Weekly

Mimic runs a video generation model as its robot policy at Audi

Lukas Ziegler got the stack explained to him at NVIDIA GTC in Berlin, and the headline detail is where the training compute went.

Mimic's policy is a video generation model with a thin action decoder bolted on, and it is running on a factory floor. Lukas Ziegler ran into the company at NVIDIA GTC in Berlin and posted the breakdown. The policy is called FLUX-mimic, built with Black Forest Labs on the FLUX 3 backbone, and Ziegler says it is running at Audi on kitting, ECU insertion and soft parts like seals and cables.

Lukas Ziegler
@lukas_m_ziegler
X
Over 95% of training compute went into predicting video. Actions are a thin readout on top.
Oct 26, 2026 · View on X

The number that matters is the split. Over 95% of training compute went into predicting video. The action decoder is lightweight and reads intermediate features from the backbone's video prediction path. Actions are a thin readout on top of a model whose actual job is guessing what the next frames look like.

The backbone does not even need finetuning

It works with the backbone frozen. Ziegler says Black Forest Labs reports the decoder beating prior VLAs (models that turn camera images and text into robot actions) with no backbone finetuning at all, a setting where VLAs reportedly fail outright. His read is that the manipulation-relevant structure is already sitting in the video representation. Nobody taught the model contact or cause and effect as a separate objective. It fell out of predicting tens of millions of hours of video.

The hardware around it is Mimic's own. Five-finger hands built in house, two wrist cameras plus a workspace camera, and their own whole-body stack doing impedance control on the arms.

Takeovers are the data pipeline

The detail Ziegler flags from the demo is drop-in teleop (a human taking control remotely). The operator steps in mid-episode to correct a wrong move, then hands back and the policy resumes. That is the safety story, and it is also a labelling machine. Every correction is a labelled intervention at precisely the moment the policy was wrong, which is the exact data a policy needs most and the hardest kind to collect on purpose.

This follows the GTC run itself, where Mimic's Elvis Nava said the hand ran fully autonomously nonstop for two days, earlier reported. Chris Paxton, who saw it, called it impressive and described it as a dexterous end to end manipulation policy running non stop.

Shown versus claimed. The GTC demo is what people watched. The Audi deployment on kitting, ECU insertion and soft parts comes from Ziegler relaying what Nava explained to him, and no video of the Audi line is in the sources. The frozen-backbone benchmark result is attributed to Black Forest Labs, not independently checked.

Elvis Nava
@elvisnavah
X
We've been running a fully autonomous demo of FLUX-mimic nonstop at GTC for two days.
Oct 22, 2026 · View on X
Chris Paxton
@chris_j_paxton
X
Running dexterous e2e manipulation policy non stop. Impressive
Oct 23, 2026 · View on X

Sound like you were at the demo

One email a day. Under three minutes. The demos, the deployments and the part the press release left out.

Get the next one by email

Robotics every day from the people building it. The demos, the deployments and the arguments worth your time. Every claim links back to the engineer.