Research

Questions we're working on

Cogninoid Labs pursues research at the frontier of physical AI — where machines must not just reason, but act, sense, and adapt in the real world.

Research Areas

Embodied AI & Learning

Studying how AI systems can learn from physical interaction with the world — through reward signals, demonstrations, and closed-loop feedback rather than purely offline data.

Robotic System Design

Designing robot hardware and software architectures that are modular, reusable, and amenable to AI-driven control across a range of manipulation and locomotion tasks.

Agentic AI Systems

Building AI agents that can plan, write code, call tools, and execute multi-step scientific workflows autonomously — with human oversight as needed.

Autonomous Experimentation

Developing closed-loop platforms where AI reasons over experimental data and proposes next steps — accelerating the scientific discovery cycle.

Materials & Scientific AI

Applying machine learning to inverse design, property prediction, and synthesis planning for advanced materials — including MXenes, perovskites, and nanomaterials.

Human–Machine Interfaces

Researching how to design interfaces that give humans interpretable, actionable visibility into AI behavior — especially in physical robotics contexts.

Philosophy

How we do research

Build systems, not isolated demos
Connect AI models with real-world physical feedback
Keep human control and interpretability central at all stages
Use rapid prototyping to shorten the idea-to-experiment cycle
Prefer open, modular, and reusable tools over closed solutions
Validate ideas in hardware as early as possible

Experiment-first mindset

We prioritize building and testing physical systems early — because real-world behavior always surprises what simulation or theory predicts.

Every research thread at Cogninoid Labs starts with a question and ends with something you can touch, run, or measure.

Notes & Thinking

Recent notes

Embodied AIMay 2026

On closed-loop intelligence

Notes on why static training datasets are insufficient for physical AI — and why the feedback loop between sensing, reasoning, and actuation is the core problem.

RoboticsApril 2026

Designing robotic hardware for AI control

Reflections on building robot arms and grippers that are explicitly designed for AI-driven control — not just traditional position-control teleoperation.

AI AgentsMarch 2026

Agentic systems for scientific workflows

Early experiments with LLM-based agents that can plan and execute multi-step lab-style workflows — with tool use and error recovery.