Flex-π: a 6B world-action model beats π0.5 by up to 6x on real bimanual tasks
chris_j_paxton · x · 2026-09-23
Researchers from the University of Washington and Allen AI released Flex-π, a 6B-parameter world-action model that jointly denoises RGB, 3D geometry, object-centric DINO semantics and actions in a shared latent space. Per-stream dropout yields one checkpoint that runs on any subset of streams, letting developers pick a speed-accuracy operating point at deployment. On a real bimanual YAM workcell — contact-rich, sub-millimeter and long-horizon tasks including gripper self-repair — Flex-π averages 83% task completion in-distribution, beating the strongest baselines by up to 2–6x while running faster than π0.5. Paper and code are public; a RoboPapers podcast episode is coming soon.
More from Embodied
- rei_labs unveils Adapt-1 Machina: RL learns continuous control sequences without demos or critic — burny_tech · 2026-09-23
- Midcentury Exits Stealth With $15M Seed to Build Data Infra for Physical AI — Scobleizer · 2026-09-23
- Karma: an open-source robot control stack for VR teleop and policy inference — k7agar · 2026-09-23
- Apple reportedly building screenless fitness tracker to rival Whoop, launch as early as 2028 — Polymarket · 2026-09-23
- Toyota to Spend $6.42B a Year on 400,000 Factory Robots From 2028 — Ars Technica AI · 2026-09-23
- Musk announces Grok bot is now in Tesla — vertigoruntime · 2026-09-23