X-Era AI Lab's world-action model hits 85-88% robot manipulation success rates

jiqizhixin · x · 2026-11-14

X-Era AI Lab released a preprint for E0.5, a world-action model built directly on trainable 3D foundation models:

Core idea: Instead of using geometry only as an auxiliary input to a video model, geometric representations are central to generating robot actions — a language-conditioned 3D geometric backbone (DINO + T5 encoders, 24 visual-frame attention blocks) feeding 24 geometry-conditioned DiT action blocks with flow matching.
Joint learning: The model learns action chunks alongside future depth, point maps, and camera parameters aligned to the chunk's final step; future geometry supervises training, and no explicit future scene generation is needed at inference.
Layerwise geometry-action attention: Action tokens attend to visual tokens one-way, no reverse attention.
Results: After embodied pretraining, E0.5 reports 85.0% average success on AgileX Piper and 88.3% on Zerith H1 across in-distribution and out-of-distribution settings (specific manipulation tasks and training regimes, not arbitrary behavior).

More from Embodied