Runway released Praxis-1 as open weights that turn video pretraining into robot control. Runway describes Praxis-1 as a world action model that predicts short video chunks and the motions to match them.
Praxis-1 scored 46.2% on RoboLab-120, ahead of FLUX 3 Action at 42.9%, according to the RoboLab-120 leaderboard cited by Runway.
Praxis-1 couples a 26.9B parameter video backbone with a 16 layer action expert, Runway says. Each call returns 24 action rows covering 1.6 seconds at 15 Hz, Runway says. Only about 2.8B backbone parameters are active per token, and the full 27.2B stack fits in 54.4GB in bf16 on one 80GB GPU, Runway says.
Praxis-1 learned in four stages starting from Runway video pretraining and moving to chunk-causal embodied video, Runway says. Mid-training used about 58% egocentric human video, 20% general video and 18% real robot video, Runway says. Fine-tuning for the Franka arm used only successful DROID episodes totaling about 277 hours, Runway says.
Praxis-1 drives both a bimanual YAM and a single Franka through contact tasks like placing a mug upside down and closing a drawer, Runway says.
Praxis-1 also adapts to unseen embodiments, Runway says. The same shared backbone and expert control Google RT-1 in SimplerEnv and fine-tune to the low cost SO-101, Runway says. That swappable head design matters because most robot labs cannot afford a separate 27B train per arm.
StartupHub.ai data groups Runway with generative video peers and NVIDIA with robotics foundation models, and Praxis-1 blurs that boundary. NVIDIA describes GR00T 2 as jointly predicting future visual observations and embodied actions across grippers, dexterous hands and humanoids, per its research lab. Praxis-1 still needs data center class GPUs and clean context chunks, which limits shop floor use, Runway says.
Can video pretraining alone carry Praxis-1 from tray and drawer demos to noisy warehouses without far more robot data?
