InternW0: A Foundational Physical World Model for Efficient Real-World Interactions
Abstract Overview
InternW0 is the first model in the InternW physical world model series, designed to connect future visual prediction with continuous robot control for closed-loop interaction in the physical world. Its architecture combines a high-capacity video expert and a lightweight action expert in an asymmetric video-action mixture-of-transformers, with observation-conditioned context routing so that action updates can use cached predictive context while incorporating new observations. The model is pretrained on about 7,200 hours of heterogeneous robot and egocentric data, including the 275-hour EgoLab wet-lab egocentric dataset, and can be extended with force and tactile histories for contact-aware manipulation. The paper evaluates InternW0 on simulation benchmarks, real-robot tasks, wet-lab procedures, and inference-efficiency measurements.
Novelty
The paper’s main novelty is an asynchronous duplex world-model design that separates slow visual prediction from fast action generation while linking them through observation-conditioned editing of cached video context. It also combines cross-embodiment pretraining, video-only egocentric supervision, and contact-aware post-training that jointly predicts actions and future interaction signals such as force and tactile feedback.
Results
InternW0 achieves 98.6% average success on LIBERO, 93.12% average success on RoboTwin 2.0-Full, and 75.60% average success on RoboTwin 2.0-Clean2Random, which is the best average reported in that comparison table. On five real-world tasks, it matches or exceeds the reported baselines, including 68.4% progress on a 15-stage metal-organic framework experiment and 65.3% progress on quantitative pipetting. Its critical-path action latency is reported as 60.73 ms, corresponding to a 16.47 Hz model-side policy update rate and a 3.13× speedup over Fast-WAM.
Key Points
- InternW0 uses an asymmetric Video-DiT/Action-DiT architecture with observation-conditioned chunk K/V editing to reuse predictive video context during fast closed-loop control.
- The training recipe spans 7,233.5 hours from seven datasets, including the 275.4-hour EgoLab egocentric wet-lab dataset and a unified 37-dimensional action/state interface for heterogeneous embodiments.
- The evaluation shows strong simulated manipulation performance, improved real-robot and wet-lab task execution, and lower latency on the action-critical inference path.
References
- arXiv: https://arxiv.org/abs/2609.27656v1
- Fugu-MT: https://fugumt.com/fugumt/paper_check/2609.27656v1
- Hugging Face Papers: https://huggingface.co/papers/2609.27656
- Project: https://internrobotics.github.io/internw0