Runway open-sources Praxis-1: 27B robot model trained on web video, tops RoboLab-120 at 46.2%
iamfakhrealam · x · 2026-11-14
Runway released Praxis-1, its first world action model, with open weights under Apache 2.0: a 27B mixture-of-experts diffusion transformer that predicts video and actions jointly, with inference and fine-tuning code (full FT and LoRA) included.
Key findings:
- Web video over robot data: robot demos are scarce but web video is "effectively limitless"; pretraining on ordinary video worked about as well as real robot footage — 16.1 cm vs 16.0 cm placement error
- Cross-scene generalization: one policy ran in a studio and a home kitchen without retraining
- Results: tops NVIDIA's RoboLab-120 with 46.2% success over 1,200 episodes, vs FLUX 3 Action (42.9%), Cosmos3 Nano (36.8%) and π0.5 (28.0%); best open-weight VLA
- Partners: Noble Machines, Standard Bots and Ultra already run it on a legged humanoid, a six-axis industrial arm and a dual-arm warehouse robot
- Availability: GitHub, Hugging Face (native LeRobot support), vLLM-Omni, hosted inference on Reactor; runs on NVIDIA hardware from data-center GPUs up
More from Embodied
- UC Davis Neuroprosthetics Lab Brings 8 Posters on Speech BCIs to SfN 2026 — neuro_engineer · 2026-11-14
- Dev data: Quest peak CCU ~10x the top 2026 SteamVR release, backing 90% market share claim — ID_R_McGregor · 2026-11-14
- Tactile model weights open-sourced in Tiny and Small sizes, links included — dr_cintas · 2026-11-14
- 4,983 hours of real touch data lift robot policy success rates across 26 tasks — dr_cintas · 2026-11-14
- Robbyant Open-Sources LingBot-Touch, Giving Robots a Sense of Touch — dr_cintas · 2026-11-14
- Neural Implant User Says AI Turns Silently Mouthed Songs Into Real Creations — Kyrannio · 2026-11-14