InternW0

A Foundational Physical World Model
for Efficient Real-World Interactions

Physical Intelligence Team, Shanghai AI Laboratory

PaperBibTeX

InternW0 overview: heterogeneous pretraining data, asynchronous video–action modeling, and simulation results — InternW0 couples future visual prediction with continuous robot control through an asymmetric video–action architecture. A high-capacity video expert provides predictive context, while a lightweight action expert adapts it to incoming observations for efficient, contact-aware interaction.

98.6%

LIBERO average

Mean success rate across all four task suites.

93.12%

RoboTwin 2.0-Full

Average success across clean and randomized settings.

75.60%

Clean2Random average

Clean-only post-training, evaluated on both scene settings.

16.47 Hz

Model-side policy updates

60.73 ms per action update on RTX 5090D; video prediction runs asynchronously.

Introduction

A Quick Overview of InternW0

See InternW0 carry out a multi-stage laboratory workflow before exploring the architecture, training data, and evaluation results.

Model

Asynchronous video–action modeling with contact-aware control

InternW0 uses a mixture-of-transformers backbone with separate video and action experts. Observation-conditioned context routing keeps longer-horizon predictions useful as new visual and physical feedback arrives.

InternW0 video and action experts coupled through observation-conditioned context routing

Data

7,233.5 hours across 25 training domains

Seven datasets combine real-robot trajectories, simulated manipulation, and egocentric laboratory video. A unified 37-dimensional state–action interface and validity masks accommodate different embodiments, while square-root domain sampling balances the training mixture.

EgoLab: egocentric laboratory activities, reconstructed hand annotations, and activity distribution — EgoLab contributes 275.4 hours of egocentric laboratory video, covering instrument handling, liquid transfer, mixing, and other wet-lab activities. Reconstructed hand trajectories support filtering; pretraining uses the video-only pathway.

Source	Type	Domains	Hours	Episodes
InternData-A1	Simulated	6	3,494.3	568,194
AgibotWorld	Real	2	2,620.0	158,380
RoboCOIN	Real	13	438.1	59,952
Galaxea	Real	1	314.9	15,362
MolmoAct	Real	1	70.3	3,424
RoboDojo	Simulated	1	20.5	3,465

Robot manipulation data mixture. Six robot datasets supply state–action supervision; domain counts distinguish data sources and configurations rather than unique robot morphologies.

Source	Hours	Episodes	Supervision
EgoLab	275.4	3,192	Future-video prediction without robot action labels

Egocentric video supervision from EgoLab. Together with the robot datasets, the full mixture contains 811,969 episodes across 25 training domains.

Benchmarks

Strong single-arm manipulation and bimanual generalization

LIBERO evaluates 40 single-arm tasks across spatial, object, goal, and long-horizon suites. RoboTwin 2.0 covers 50 bimanual tasks, with Full and Clean2Random protocols separating in-domain adaptation from generalization to unseen randomized scenes.

Benchmark	Metric	InternW0
	Average success rate	98.6%
	Average success rate	93.12%
	Average success rate	75.60%

Success rates are reported in percent. Click a benchmark row to view the full comparison and evaluation setting.

Limited-data RoboTwin Clean2Random ablation comparing no pretraining, robot-only pretraining, and robot plus EgoLab — In a limited-data ablation using about 50 hours of robot data, adding about 50 hours of EgoLab improves randomized-scene success from 13.73% to 21.97% (+8.24 percentage points). These ablation results use a smaller corpus than the main Clean2Random result.

Critical-path latency and model-side update rate: Motus, Fast-WAM, and InternW0 — InternW0 takes 60.73 ms per critical-path action update on RTX 5090D, corresponding to a maximum model-side update rate of 16.47 Hz and a 3.13× speedup over Fast-WAM. This measurement includes observation encoding, context routing, and action denoising; asynchronous video-plan generation is excluded.

Experiments

From everyday manipulation to scientific workflows

Five real-world tasks test instruction following, object discrimination, long-horizon coordination, and contact-aware tool use. Four use a dual-arm platform; quantitative pipetting uses a 20-DoF dexterous hand on a 7-DoF arm with hybrid force–position control.

Five real-world tasks: Make Sandwich, Pick Industrial Parts, Sort Tubes, MoF Experiments, and Quantitative Pipetting — The real-world suite spans Make Sandwich, Pick Industrial Parts, Sort Tubes, MoF Experiments, and Quantitative Pipetting.
Make Sandwich
Pick up two bread slices and a piece of meat, then stack them in the prescribed order.
Pick Industrial Parts
Identify industrial parts and place each one into its matching box.
Sort Tubes
Follow an instruction specifying the arm, tube color, and destination box.
MoF Experiments
Complete 15 ordered subtasks for solution preparation, including pouring, funnel removal, flask transfer, stopper insertion, and stirring.
Quantitative Pipetting
Coordinate five stages of dexterous tool use: pickup and reorientation, tip attachment, aspiration, dispensing, and tip ejection and return.

Real-world results

Method	Make Sandwich · SR	Pick Industrial Parts · SR	Sort Tubes · SR	MoF · Progress	Pipetting · Progress
π0.5	73.3	53.1	86.7	50.2	46.7
Fast-WAM	40.0	50.6	66.7	10.7	18.7
InternW0	73.3	82.7	88.9	68.4	65.3

Results over 15 trials per task with randomized initial object poses (Table 5). Sandwich uses episode success rate; parts and tubes use object-level success rate. MoF and pipetting use ordered-subtask progress rate, with the same VLM for subtask generation across methods. All values are percentages.

Everyday and industrial manipulation

Dual-arm demonstrations of ordered assembly, part sorting, and language-conditioned tube selection.

Make Sandwich

Pick Industrial Parts

Sort Tubes

MoF Experiments

Close-up views of flask placement, stopper insertion, and stirrer operation within the 15-stage solution-preparation workflow.

Flask placement

Stopper insertion & stirring

Quantitative Pipetting

A dexterous hand coordinates pipette pickup, tip attachment, aspiration, dispensing, and tip ejection with force-aware control.

Quantitative Pipetting

MoF Experiments — subtask results

Subtask	π0.5	Fast-WAM	InternW0
1 · Pick up funnel from rack	100.0	100.0	100.0
2 · Insert funnel into flask	100.0	46.7	100.0
3 · Pick up graduated cylinder	100.0	13.3	93.3
4 · Pour liquid into flask via funnel	73.3	0.0	80.0
5 · Place graduated cylinder back	73.3	0.0	80.0
6 · Hold flask steady on platform	73.3	0.0	80.0
7 · Pick up funnel from flask	73.3	0.0	80.0
8 · Insert funnel into rack hole	46.7	0.0	66.7
9 · Return to initial position	46.7	0.0	66.7
10 · Place flask onto stirrer	33.3	0.0	66.7
11 · Pick up stopper from rack	6.7	0.0	66.7
12 · Insert stopper into flask	6.7	0.0	66.7
13 · Press left button of stirrer	6.7	0.0	26.7
14 · Press right button of stirrer	6.7	0.0	26.7
15 · Return to initial position	6.7	0.0	26.7
Overall progress rate	50.2	10.7	68.4

Subtask success rates and overall progress rate (%) from Table 6. Overall progress is the fraction of subtasks completed correctly and in order.

Quantitative Pipetting — subtask results

Subtask	π0.5	Fast-WAM	InternW0
1 · Pickup & reorientation	93.3	46.7	86.7
2 · Tip attachment	40.0	13.3	66.7
3 · Liquid aspiration	33.3	13.3	66.7
4 · Liquid dispensing	33.3	13.3	60.0
5 · Tip ejection & return	33.3	6.7	46.7
Overall progress rate	46.7	18.7	65.3

Subtask success rates and overall progress rate (%) from Table 7. Contact-aware post-training combines force history with visual and proprioceptive observations.

Pouring and pipette tip-attachment examples comparing InternW0, pi 0.5, and Fast-WAM — Qualitative comparisons on pouring (top) and pipette tip attachment (bottom). InternW0 aligns the tool before contact-rich execution and uses force feedback to adapt contact, while the illustrated baseline rollouts show misalignment.

Citation

BibTeX

@misc{cai2026internw0foundationalphysicalworld,
  title         = {InternW0: A Foundational Physical World Model for Efficient Real-World Interactions},
  author        = {Jisong Cai and Yao Mu and Ganlin Yang and Zhe Cao and Zhangzheng Tu and Xing Gao
                   and Kailin Li and Xinyu Zhan and Lixin Yang and Yangkun Zhu and Haoxiang Ma and Ming Zhou
                   and Qiaojun Yu and Yufei Xue and Liqun He and Yifei Yao and Yifan Zhu and Long Ling
                   and Bingqi Jiang and Haoyu Guo and Xueyue Zhu and Bowen Zhou and Bin Zhao
                   and Tianfan Xue and Chunhua Shen and Weinan Zhang},
  year          = {2026},
  eprint        = {2609.27656},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2609.27656}
}

InternW0

LIBERO average

RoboTwin 2.0-Full

Clean2Random average

Model-side policy updates

A Quick Overview of InternW0

Asynchronous video–action modeling with contact-aware control

7,233.5 hours across 25 training domains

Strong single-arm manipulation and bimanual generalization

Egocentric video improves transfer

Efficient action updates

From everyday manipulation to scientific workflows

Real-world results

Everyday and industrial manipulation

MoF Experiments

Quantitative Pipetting

Contact-aware manipulation

BibTeX