Language Models
DescribeRepresent concepts through language and support semantic reasoning. Physical state remains indirect rather than a persistent representation for control.
Physical AI · Concept World Model
Teaching machines the logic behind reality.
Understand the physical world. Act reliably. Evolve continuously.
Awomo builds a concept-first, world-model-driven physical intelligence core that perceives multimodal environments, represents physical state and constraints, predicts action outcomes, and generalises across scenes, tasks and hardware embodiments.
Concept-firstPredict · decide · actOne understanding · many embodiments
Object stateposition · geometry · affordance
Physical constraintcontact · friction · collision
Action hypothesisintervene · simulate · compare
Predicted outcomerisk · motion · consequence
Technology
A different foundation for physical intelligence.
Language models, video-generation models and fixed simulations each serve a distinct purpose. Awomo's Concept World Model is designed around the states, constraints and consequences required for reliable action in the physical world.
Existing approaches
What they primarily represent
Each approach is valuable within its intended domain, but it is optimised for a different representation and objective.
Tokens and semantic relationships
Video Generation
VisualisePredicts or generates how a scene may look. Visual plausibility alone does not guarantee consistent contact, force, collision or controllability.
Fixed Simulation
ReproduceReproduces behaviour under explicitly specified models. Accuracy depends on how completely the environment and long-tail conditions are defined.
Primary outcomes
DescribeVisualiseReproduce
Awomo's approach
Concept World Model
An action-relevant model of the physical world, built for prediction, decision and control.
State
Constraint
Action
Outcome
Physical
Understanding
Understanding
01
Objects and relationshipsWhat exists and how entities are connected.
02
Physical states and constraintsWhat can move, interact or change.
03
Actions and possible outcomesWhat may happen under an intervention.
04
Risk-aware decision-makingWhich action best satisfies the goal and constraints.
05
Feedback and continuous learningHow the model improves after real-world deployment.
Primary outcomes
UnderstandPredictActImprove
Not another combination of existing models.
A different representation objective, designed for interaction with the physical world.
Perceive. Represent. Understand. Predict. Act. Improve.
A connected Concept World Model path turns multimodal observation into physical reasoning, action control and continuous model improvement.
Concept World Model
Technical Path
01
Multimodal Physical Perception
Fuse images, point clouds, audio, sensors and action feedback.
02
Latent Space Representation
Compress complex physical phenomena into an efficient latent space.
03
Concept World Model
Represent objects, relationships, states, actions and goals.
04
Prediction & Action Control
Simulate possible outcomes and evaluate risk before acting.
05
Self-Evolving Loop
Discover failures, diagnose causes, add targeted data and update the model.
Advantages
Learn more efficiently. Generalise more reliably.
The Concept World Model focuses computation on action-relevant physical concepts and causal structure, rather than scaling pixels alone.
High data efficiency
Learn the physical concepts and causal structures that matter, instead of relying only on pixel-level accumulation.
Low inference cost
Reduce high-resolution generation overhead for real-time control, edge deployment and responsive decision-making.
Strong generalisation
Reuse one physical-intelligence core across tasks, environments and different hardware embodiments.
Technology · Validation
Validation and demo
StrucPhysVideo Achieves 45.5 on Physics-IQ with 30B-A3B Model
Awomo introduces StrucPhysVideo, a family of video world models that combines physics-focused data curation with language- and action-conditioned prediction of how scenes evolve.
Read the full articleApplications
Physical intelligence for complex real-world operations.
Autonomous Systems
Autonomous driving
Model long-tail traffic scenarios, anticipate risk and support policy validation across changing road conditions.
Designed for weather, roadworks, occlusion and rare interactions where early prediction is critical.
Embodied Intelligence
General robotics
Enable manipulators, mobile robots and humanoids to reason about affordances, spatial constraints and action consequences.
Support reliable task transfer across objects, environments and hardware embodiments.
Industrial Intelligence
Industrial manufacturing and logistics
Support adaptive automation in production, warehousing, inspection and specialist operations.
Maintain performance when layouts, object specifications and workflows change.
Data Flywheel
Simulation and data generation
Identify failure modes, diagnose root causes and generate or select targeted training data.
Focus data generation on measurable model gaps rather than undirected scale.
Model Integration
Intelligent-hardware model licensing
Embed Awomo’s physical-intelligence capabilities into robots, connected devices and industrial hardware.
Integrate through licensed models, runtimes or APIs without rebuilding the intelligence stack from the ground up.
Deployment
One model. Multiple ways to deploy.
Awomo provides the physical-intelligence core and works with customers and partners to match the delivery model to the task, infrastructure and hardware.
CWM
One understanding, many embodiments.
Adapt the same physical-intelligence foundation to robots, vehicles and intelligent devices while respecting project-specific hardware and data boundaries.
Integrated delivery
Full-system solutions
A task-specific solution combining Awomo models, selected hardware, sensors and integration support.
- Defined operating scope
- Hardware and sensor integration
- Pilot, validation and deployment
Controlled infrastructure
Private deployment
Dedicated cloud or on-premises deployment for strict data, security or latency requirements.
- Customer-controlled data boundary
- Project-specific adaptation
- Security and retention controls
Software integration
Model and API licensing
Licensed models, APIs and SDKs for adding physical-state modelling, prediction and decision support to existing products.
- API and SDK integration
- Model licensing
- Runtime adaptation
Partner ecosystem
Global Hardware Partners
Compatibility and joint integration with robot, sensor and computing-platform partners across major markets.
- Platform compatibility
- Joint technical validation
- Regional deployment support
About Awomo
Originating from AutoLab, building foundation models for the physical world.
Awomo focuses on Physical AI, combining original world-model research with large-scale engineering and real-world deployment.
2026Company founded
20+PhDs and senior engineering experts
USD 15M+Raised to date
China + SwitzerlandResearch, engineering and partnership presence
Brand Definition
A Physical AI technology provider for the real world.
Awomo develops Concept World Models for embodied AI, autonomous systems, industrial intelligence, intelligent hardware and simulation—turning physical understanding into reliable action.
Not a million repetitions. One concept, understood well enough to generalise. Awomo aims to move intelligence beyond text, images and screens—into real devices, tasks and environments.
A
Autonomous
Autonomous adaptation, decision-making and continuous evolution.
WO
World
Grounded in the structures, constraints and dynamics of the physical world.
MO
Model
A foundation world model that understands, acts and improves.
The core team includes Westlake University faculty and more than 20 PhDs and senior engineering experts. The team collaborates closely with Westlake University and the University of Chinese Academy of Sciences across world models, multimodal models, reinforcement learning, simulation, data systems, autonomous driving and engineering infrastructure.
Meet the team
Meet the team.
Founder
Kaicheng YU
Kaicheng holds a PhD from EPFL and is an Assistant Professor and Head of the Autonomous Intelligence Laboratory at Westlake University. He has led research in autonomous driving, multimodal AI, and embodied intelligence. As Founder of Awomo, he is building Concept World Models that enable machines to understand and act in the physical world.
Research Partner
Tong ZHANG
Tong holds a PhD from ANU and completed postdoctoral research at EPFL. He is a tenure-track Assistant Professor at UCAS, as well as a Host Professor at EPFL’s IC IINFCOM IVRL Lab. His research focuses on world models, autonomous driving, spatial intelligence, representation learning, and 3D vision. As Research Partner of Awomo, Tong leads the company’s core algorithm development and technical strategy.
Co-founder
Tim XU
Tim is a senior expert in multimodal AI and foundation models, with extensive experience in end-to-end model training, deployment, and industrialisation. He has led algorithm teams supporting products used by hundreds of millions of users, published multiple research papers, and holds more than ten patents. As Co-founder at Awomo, he leads multimodal model development and engineering deployment.
Research Partner
Tao LIN
Tao holds a PhD and an MSc from EPFL. He is a tenure-track Assistant Professor at Westlake University and leads LINs Lab. His research focuses on efficient and robust deep learning, optimisation, generalisation, and collaborative learning. As a Research Partner at Awomo, he contributes to video representation learning and pre-training and leads efforts to accelerate large-scale model training.
Co-founder
Yao DI
Yao holds degrees from EPFL, SJTU, and HKU, with a multidisciplinary background in technology management, engineering, and mathematics. Yao brings over ten years of experience in R&D operations, strategy, compliance, and risk management. As Co-founder of Awomo, Yao leads corporate functions and supports the company’s operational development.
Business Development Manager
Xiaofeng LI
Xiaofeng holds an MS in Electrical Engineering from NYU and has a multidisciplinary background in AI, engineering, and international collaboration. As Business Development Manager at Awomo, Xiaofeng supports technical communication, project coordination, and cross-functional collaboration.
Join the team
Bring foundation models into the real world.
Physical intelligence is still at an early stage. We are looking for people who understand models, infrastructure, simulation, control and real environments.
Foundation Models — World Model↗
Post-training↗
Training and Inference Infrastructure↗
Reinforcement-Learning Simulation↗
Data Algorithms and Pipelines↗
Robotics Deployment / Sim2Real↗
IT Infrastructure and Operations↗
Location: Hangzhou, Beijing, Switzerland · Applications: info@awomo.ch · Suggested subject: Name + Position
News
Latest from Awomo.
StrucPhysVideo Achieves 45.5 on Physics-IQ with 30B-A3B Model
Awomo introduces StrucPhysVideo, a family of video world models that combines physics-focused data curation with language- and action-conditioned prediction of how scenes evolve.
Read articleAwomo Secures Over USD 15 Million in Funding to Accelerate the Development of Physical AI Infrastructure
Awomo has completed a series of seed and angel financing rounds, jointly backed by Innoangel Fund, Dongfang Jiafu, Zhengxuan Investment, Tianqi Capital, the Westlake Innovation Investment Fund and Jinma Investment.
The funding will further develop the data, computing and model infrastructure required for Physical AI, alongside product development and validation in industrial automation, robotic manipulation and digital twins.
Read the full announcementFAQ
A few common questions.
Awomo builds Concept World Models that give robots, vehicles and intelligent hardware foundation-model capabilities for understanding the physical world and acting reliably within it.
A Concept World Model represents physical state, spatial relationships, motion and causal change in latent space. It focuses on the physical understanding required for action rather than only generating the next frame.
Different hardware embodiments operate in the same physical world. Awomo aims to transfer learned physical regularities across vehicles, robot arms, humanoids and intelligent devices, reducing repeated development and adaptation cost.
Awomo primarily develops the Concept World Model, runtime and physical-intelligence software stack. For selected deployments, complete task-oriented systems can be delivered with hardware and integration partners.
Depending on the project, Awomo can support model or API licensing, private deployment, and partner-delivered full-system solutions. The final architecture depends on the task, hardware, data boundary and success criteria.
Awomo is developing a Swiss-based European operation and can explore private infrastructure and EU-region data-residency options. Specific GDPR, residency and security claims depend on the agreed architecture, subprocessors and contract.
Contact us
Start a Physical AI pilot.
Tell us about the physical task, target hardware, data boundary and success criteria. We will use this information to scope the most appropriate collaboration and deployment model.
Emailinfo@awomo.ch