Introducing Flex-π: A 6B World-Action Model Bridging VLAs and WAMs

LuizaJarovsky · x · 2026-08-13

Researchers introduced Flex-π, a 6B-parameter multi-stream World-Action Model (WAM). It jointly predicts future RGB, 3D pointmaps, and DINO semantics alongside actions during training. At deployment, it can flexibly run as a VLA, a full WAM, or anything in between from a single checkpoint.

Experiments show that Flex-π outperforms the strongest baselines by 2–7× on dexterous, precise, real-world bimanual tasks while running faster than π0.5, demonstrating strong in- and out-of-distribution generalization.

Related event: Flex-π: 6B Parameter Robotic Model Released(2 posts)→

Original post →

More from Embodied

Embodied channel →