Research
Brain-inspired AI for robots that perceive, remember and reason
We research cognitive architectures, memory, perception and simulation for autonomous systems. The goal is machines that build understanding over time, the way people do, rather than starting from zero with every task.
01
Brain-inspired cognitive architecture
Most AI systems are a single model doing one job. Brains are different: specialized systems for sensing, remembering, attending and planning, coordinated in real time. We study cognitive architectures that borrow that structure, so machines can combine fast reactive behavior with slower, deliberate reasoning.
Questions we are exploring
- How should fast reflexes and slow deliberation share control of one robot?
- What role should attention play in deciding what an agent notices and stores?
- Which parts of human cognition are worth copying, and which are not?
02 · Main research direction
Memory for autonomous agents
Today’s AI forgets almost everything between tasks. We are developing a brain-inspired approach to associative memory, where knowledge is stored as connected units that link, strengthen and fade with use, so an agent can recall the right experience at the right moment. This is our main research direction; details will follow publication.
Questions we are exploring
- How can an agent decide what is worth remembering?
- How should memories connect, strengthen and fade over time?
- How do episodic memories (what happened) become general knowledge?
03
Perception for robots
A robot is only as capable as what it can perceive. We explore multimodal perception that fuses vision, depth, touch and audio into a shared picture of the world, and grounds language in what the robot actually senses, so instructions like “check the valve by the pump” mean something concrete.
Questions we are exploring
- How do we fuse vision, depth, touch and sound into one world model?
- How can a robot actively move to see what it needs, instead of waiting?
- How do we ground language in perception reliably enough for real work?
04
Simulation and digital twins
Real-world robot data is slow and expensive to collect, and mistakes are costly. Simulation lets us train and stress-test agents across thousands of scenarios, generate synthetic data, and build digital twins of real facilities. The hard part is the gap between simulation and reality, and that is where we focus.
Questions we are exploring
- How do we close the sim-to-real gap for perception and control?
- Can digital twins of real buildings make robot deployment faster and safer?
- Which failures can simulation catch before they happen on site?
05
Embodied learning and control
Intelligence in the physical world has to be learned through a body. We look at learning from demonstration, reinforcement learning and foundation models for robot control, with safety constraints built in from the start rather than added afterwards.
Questions we are exploring
- How few demonstrations does a robot need to learn a new task?
- How do we keep learning safe when a robot is practicing in the real world?
- Where do large robot foundation models help, and where do they fall short?
06
Evaluation, safety and trust
Claims about AI are easy; evidence is harder. We design evaluations that measure capability honestly, study interpretability so people can see why a system acted, and connect research to the governance practices organizations need before autonomous systems go to work.
Questions we are exploring
- What should a fair benchmark for robot cognition measure?
- How can operators understand and correct an agent’s decisions?
- What governance does an organization need before deploying autonomy?
How we do research
Learn from the brain, not imitate it
We borrow structure from neuroscience and cognitive science where it helps machines work better, not for its own sake.
Built for the physical world
Ideas are judged by whether they help a real robot perceive, remember and act in real buildings.
Evidence before claims
Results are shared once they are measured and reproducible. Until then, we describe directions and open questions.
Safety by design
Evaluation, interpretability and governance are part of the research, not an afterthought.
Current work
Associative memory for autonomous agents
A brain-inspired memory system in which knowledge units connect, strengthen and fade with use. A preprint is in preparation.
Collaborate with us
We welcome research institutions, robotics teams, investors and grant programs interested in brain-inspired AI, robot memory, perception and simulation.