InternW0-Δ is a World Action Model for generalist robot manipulation that combines pretrained visual dynamics, scene semantics, 4D geometric and motion priors, and action generation. It uses a Mixture-of-Transformers architecture where a video expert and an action expert interact under guidance from a frozen vision-language model, with geometric priors injected through training-only distillation. Key contributions include Causal Imprint, which learns future-relevant scene changes from future supervision without needing future video at inference, and a large open-source training corpus of over 20,000 hours of data. The model outperforms prior methods across simulation benchmarks and real-robot platforms.
Predictive visual models improve action generation for generalist robot manipulation
InternW0-Delta combines video, semantic, geometric and action models, trained on more than 20,000 hours of heterogeneous demonstrations.
Academic
Xingyu Miao · Zizun Li · Baole Fang · Kaiwen Song · Tenghui Wang · Hanxue Zhang · +42 more
Shanghai AI Laboratory
Research Digest··2 min read
In one line
InternW0-Δ jointly models visual dynamics and action generation for robot manipulation, outperforming prior methods across benchmarks.
Before you read this
Three ideas this paper builds onworld model for agentsA predictive environment model that agents use to simulate outcomes, enabling planning, reasoning, and long-horizon task execution.113 papers →vision-language-action modelA robot policy model that maps visual observations and language instructions to actions, trained on demonstrations and often used for manipulation tasks.84 papers →mixture of expertsA neural architecture with many specialized sub-networks, only a few activated per input via a router, balancing capacity and efficiency.84 papers →
Read it as
The authors built a world action model that uses pretrained visual dynamics and geometric knowledge to generate robot actions without rolling out future video during deployment.
The authors assembled more than 20,000 hours of processed robot demonstrations, UMI recordings, egocentric human demonstrations and human-to-robot converted data.
How it works
stage by stageThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.