Physical AI research
Exploring intelligencethat acts in the physical world.
Durathink's Physical AI research explores how perception, simulation, learning, reasoning, and control can support intelligent systems that interact safely and effectively with real-world environments.
Ongoing, exploratory research. Not a commercial product.
Areas of interest
What we are looking at
Early-stage work and prototypes. These are directions under active exploration, not solved problems or shipped capabilities.
- Robot learning
- How agents acquire manipulation and locomotion skills from demonstration, interaction, and structured practice.
- Simulation
- Using simulated environments to train and stress-test behaviour before it reaches physical hardware.
- Reinforcement learning
- Reward design, sample efficiency, and the gap between simulated competence and real-world reliability.
- Multimodal perception
- Combining vision, depth, force, and language into representations an agent can act on.
- Manipulation
- Contact-rich tasks where success depends on feedback and correction rather than open-loop precision.
- Human-in-the-loop control
- Where a person stays in the loop, what they are shown, and how control is handed back safely.
- Agent planning for physical systems
- Connecting high-level reasoning to low-level control under real timing and safety constraints.
Why this sits alongside our products
The same problem, one layer down
Durathink builds agents that take an ambiguous input, decide what can be resolved automatically, and hand the rest to a person. Durathink Purchase Order Agent does that with purchase orders. Physical AI asks the same question where the consequences are physical — what a system should do on its own, and where it must stop and ask.