WORLD
ACTION
MODEL

通过观察与交互学习物理智能
使机器人能够理解现实世界并自主运行
实现跨任务、跨环境、跨本体泛化

交互驱动智能持续进化

在交互中理解物理世界

通过观察与交互学习物理智能
使机器人能够理解现实世界并自主运行
实现跨任务、跨环境、跨本体泛化

TACTILE-DRIVEN REINFORCEMENT LEARNING BEYOND SPARSE REWARDS NEW DATA (REAL INTERACTION DATA) MODEL WEIGHTS EFFICIENCY + ACCURACY CONTINUAL LEARNING Tacitle feedback provides what other reward signals cannot: dense, contact-grounded physical ground truth at the moment of interaction. 
Layer 1
Data & Training

Large-Scale Observation

Efficiency
Pre-training

Diverse robot operation videos, human demonstrations, and proprioceptive recordings — providing broad priors about objects, motion, and environmental dynamics.

Human Demonstrations
Videos
Language
Proprioception

Tactile-Augmented Simulation & Training

Efficiency

Simulation generates diverse interaction variants at scale — augmented with tactile signal modeling to better reflect real contact dynamics.
Real-world tactile data from Sentra continuously recalibrates simulation parameters, ensuring virtual training remains grounded in physical reality.

Real Interaction Data

Accuracy
Pre & Post-training

High-fidelity tactile interaction data capturing force dynamics, contact transition, and material response — the physical ground truth that vision cannot provide.

Contact Force Maps
Tactile Deformation
Material Response
Slip Detection
3D Force Signals
Layer 2
Trained Models

World Action Models

PREDICTIONS GROUNDED IN HOW THE WORLD CAN BE ACTED UPON
NOT JUST HOW IT LOOKS

Inputs

Vision
Tactile
Language
Proprioception

Outputs

Physical Future Prediction
Continuous Action Policy
Layer 3
Deployment Outcomes
ENVIRONMENT-INVARIANTMANIPULATION

In contact-rich tasks, tactile signals provide continuous feedback on force, slip, and material response — revealing actionable affordances that guide stable execution through every phase of contact.

Few-ShotCross-Task

Physical interaction follows causal laws — not visual patterns. By grounding models in tactile physics, skills learned in one context transfer to unseen objects and environments without task-specific retraining.

Cross-Embodiments

Force, compliance, and contact dynamics are the universal language of physical interaction. Models trained on one platform transfer across robot morphologies — reducing deployment cost for every new embodiment.

意图驱动的灵巧操作

机器人可以根据意图进行
感知、计划和行动
无需脚本或特定任务的编程

自适应触觉交互

毫秒级的触觉实时反馈让
灵巧操作更稳定

内置通用技能基元库

可组合的技能单元
用于模块化的跨任务扩展

开放技能生态系统

专为积累、复用与演进机器人操作
能力而构建的共享技能库

OPEN MODEL...
OPEN SKILLS...
OPEN WORLD...
Robot for Science
1
Precision Engineering
2
Scalable Services
3
Consumer Robotics
4

释放物理智能
驱动商业变革