What it is
Embodied foundation models with native 3D grounding, contact-point prediction, and a cross-embodiment VLA action space for real robots. Instead of stopping at 2D grounding and per-robot action heads, it adds contact-point prediction and native 3D grounding in pretraining, then uses one cross-embodiment action space with robot-specific masking for VLA fine-tuning.
Key results
- 122B-A10B outperforms all evaluated models on VSI-Bench
- 122B-A10B outperforms all evaluated models on MMSI
- 122B-A10B outperforms all evaluated models on RefSpatial-Bench
Manipulation-aligned grounding plus shared cross-embodiment action modeling transfers better to heterogeneous real-robot policies than generic VLM initialization.