Robot runs a video generation model as its policy: frozen FLUX beats VLAs
lukas_m_ziegler · x · 2026-10-26
The author tried mimic robotics' robot at NVIDIA GTC Berlin: in-house five-finger hands, two wrist cameras plus a workspace camera, and their own whole-body impedance-control stack.
The headline is the policy: FLUX-mimic, developed with Black Forest Labs on the FLUX 3 backbone.
- A lightweight action decoder reads intermediate features from the video prediction path; over 95% of training compute went into video prediction, with actions as a thin readout on top
- The backbone stays frozen — no finetuning at all — and BFL reports the decoder beating prior VLAs in a setting where VLAs reportedly fail outright
- Manipulation-relevant structure like contact and cause-effect wasn't taught separately; it emerged from the video representation
More from Embodied
- Humanoid robots' DRAM demand may 20x, but it's still just 0.04% of 2026 supply — zephyr_z9 · 2026-10-26
- Microduck team publishes deep dives: relocalizes within 2cm using a 64-pixel depth sensor — mishig25 · 2026-10-26
- Robot's elbow-assisted get-up move goes viral for its humanlike fluidity — iamfakhrealam · 2026-10-26
- Chinese eVTOL EHang EH216-S completes nearly 100,000 flights across 23 countries — LoganGrasby · 2026-10-26
- Toyota bets $6.4 billion a year on humanoid robots and factory automation — The Decoder · 2026-10-26
- RoboJudge: 800 human-annotated videos benchmark MLLM judges for embodied video generation — CUHK-CSE · 2026-10-26