Mimic Robotics Unveils Industrial Manipulation Model FLUX-mimic, Which Uses Video Pretraining

Mimic Robotics published research on Sept. 23, 2026, about FLUX-mimic, a model that combines video representations with robot actions. Using data from 31 tasks involving automotive manufacturing parts, comprising 55,217 episodes and 295.8 hours in total, it reported success rates of 77.5% with 2 hours of action-labeled data and 95% with 8 hours for soft-part kitting with Franka.

Mimic Robotics’ evaluation presents a training design that supplements action-labeled demonstrations, which are difficult to collect for industrial manipulation, with recordings of camera video alone. The reported rates of 77.5% and 95% are experimental results for soft-part kitting with Franka; they do not show that the same performance can be achieved in factory mass-production processes.

The focus of the research is a design that uses video recordings without robot-action labels to build a foundation for manipulation policies. Wearable recordings account for about 27.4% of all 55,217 episodes and about 26.0% of the total 295.8 hours. They are used to pretrain the video backbone without action labels, while demonstrations with action labels are used for policy training. Thus, the technical focus of the report is not only the volume of recordings, but also the use of video and action information at separate stages of training. For soft-part kitting with Franka, increasing manipulation data from 2 hours to 8 hours, a fourfold increase, raised the success rate from 77.5% to 95%. Mimic Robotics also reported that the 2-hour setup outperformed both baselines when they used all their data. However, this compares the amount of manipulation data used for policy training; it does not compare total costs, including data and computing resources required for pretraining. The 35-percentage-point improvement from video prediction loss was reported in a separate evaluation, and results from experiments using 40 or 20 rollouts cannot be taken as evidence of prolonged factory operation or mass-production performance. BMW Group is listed among the authors’ affiliations, and the research used data involving automotive manufacturing parts. However, there is no indication that the system was deployed on a BMW production line. This newspaper’s report on more than 1,700 ABB robots and next-generation car production covered BMW Group production and ABB robots in a separate story, and there is no basis for linking it to the equipment or results in this research. The 96% success rate in the proof of concept described in Google DeepMind unveils Gemini Robotics 2.5, reports 96% success rate picking automotive parts also cannot be directly compared because the tasks, robot platforms and evaluation conditions differ. The industrial implication is limited to the potential to incorporate video-only recordings into training for tasks where collecting action demonstrations is difficult. Policies were trained separately for each robot platform, Franka and Trossen; the research did not demonstrate skill transfer between different platforms. The next points to examine are application to parts and processes outside the benchmark, sustained operation on actual production lines, recovery from failures, safety, and data and computing costs that include video pretraining.

Of the data, recordings collected through teleoperation of Franka’s dexterous hand and Trossen’s gripper covered 20 tasks, 40,098 demonstrations and 218.9 hours, and included camera video and robot-action labels. Video-only recordings collected by people using wearable devices comprised 15,119 episodes and 76.8 hours, and included 11 tasks not found in the teleoperation data. Recordings for four benchmark tasks were also used to train and evaluate policies, while recordings covering the remaining 27 tasks were used to pretrain the video backbone. FLUX-mimic combines the video backbone of FLUX 3, pretrained on web and industrial-work data, with an action expert that outputs robot actions. It takes task descriptions, past video and past robot states as inputs, and is trained with a loss that jointly predicts video and actions. For action inference, it denoises the backbone once and predicts actions without generating the full video. In a separate evaluation of soft-part kitting, the company reported a 35-percentage-point increase in the success rate from adding video prediction loss.