Flex-π: a 6B world-action model that predicts 3D geometry for data-efficient robots
chris_j_paxton · x · 2026-09-23
Chris Paxton's RoboPapers episode 106 (now with an auto-generated, ChatGPT-corrected transcript) covers Flex-π, a 6B-parameter world-action model. Key points:
- Most WAMs predict only RGB latents, though 3D geometry and object semantics matter more for robots than color.
- Flex-π exploits a "free lunch": the same frozen video-generation VAE that encodes RGB also encodes 3D pointmaps nearly losslessly, with no pointmap-specific training.
- This enables joint supervision of RGB, 3D pointmaps, and DINO object features with no new sensors, pre-training, or inference latency.
- Signals are projected into a shared latent space and denoised jointly with actions in a Mixture-of-Transformers backbone, with per-stream dropout and cross-modality forcing.
- Result: far better demonstration efficiency and generalization to complex, precise long-horizon tasks.
Related event: Flex-π: 6B World Action Model Predicts 3D Geometry to Boost Data Efficiency(3 posts)→
More from Embodied
- UBTECH sells 400 U1 Pro humanoids to Yongda Auto in ~$10M deal — CyberRobooo · 2026-11-12
- Baidu World 2026 to showcase Apollo Go robotaxi rides and Apollo World physical AI model — Baidu_Inc · 2026-11-12
- NVIDIA's open-source Alpamayo 2 Edge VLA driving model lands in Autoware — drmapavone · 2026-11-12
- A closer look at Astra's robot gripper design (and a joke at its expense) — _Stocko_ · 2026-11-12
- Robot startup passes first presale milestone with 10+ units claimed — R4nsack · 2026-11-12
- Runway launches Praxis-1, a 27B MoE world action model, with a 73ms Fast variant — ChongZzZhang · 2026-11-12