InternW0-Δ: A World Action Model Trained on 20K+ Hours of Open Robot Data, Fully Open-Sourced
Xingyu Miao · hf · 2026-09-28
InternW0-Δ is a unified World Action Model for generalist robot manipulation, pretrained on a heterogeneous corpus of 20K+ hours — reportedly the largest open corpus of its kind.
- Architecture: Mixture-of-Transformers with a pretrained video expert and action expert interacting under semantic guidance from a frozen VLM; a pretrained 4D foundation model injects geometric and motion priors via training-only distillation
- Causal Imprint: learns future-relevant scene changes from training-only future supervision, giving the action expert predictive representations without future-video rollout at inference
- Data: robot demonstrations, UMI data, egocentric human demos, and Ego2Robot data aligned under a common state-action representation
Outperforms prior methods on simulation benchmarks and real robots; training code, weights, infra, and data pipeline will be open-sourced.
More from Embodied
- InternW0-Δ open-sources code, weights and 20K+ hours of robot data — arankomatsuzaki · 2026-09-28
- IMLE-VLA replaces flow matching with single-step cIMLE, 3.67x faster robot actions at 55Hz — petitegeek · 2026-09-28
- Yacine demos general sim2real pipeline: policy trained in just two minutes — yacineMTB · 2026-09-28
- Only 7 Years: From Kicked Robot Dogs to a Robotic Leap — alvelda · 2026-09-28
- Berkeley's Morphometric Imitation Hits 89.3% Zero-Shot Real-World Success Across 3 Robot Hands — Berkeley · 2026-09-28
- SolidBot moves real steel: post-processed robot programs now heading into TCP and accuracy tests — MatthewChang · 2026-09-28