Connected Robotics studies robotic and sensor systems in which intelligence emerges from the communication, memory, reasoning, and coordinated action of distributed physical agents.
The problem
Devices are connected. Reasoning is not.
The question
How can connected physical agents think together?
How can connected physical agents collaboratively construct, update, and act upon shared models of the real world under partial observability, uncertainty, limited communication, and edge-computing constraints?
01/The architecture
Distributed agents contribute semantic evidence to a shared world model; reasoning over that model produces explainable collective action.
A robot team and the sensors around it can only reason about what they can collectively observe. This pillar treats coverage, occlusion, and sensor placement as decisions the system makes and revises at run time — deploying sensors, moving to see, and choosing what is worth measuring next.
Agents cannot reason together over raw sensor streams. They need a common, structured account of the world: what is present, how it relates, when it was seen, by whom, and how sure anyone is. This pillar builds semantic scene graphs and physical knowledge graphs that are dynamic, multimodal, and explicitly uncertain.
This is the centre of the program. Agents observe asynchronously, partially, and sometimes wrongly. Distributed physical reasoning asks how evidence — not raw data — propagates through a team, how beliefs are updated under uncertainty, how contradictions are detected rather than averaged away, and how consistency is maintained without a central authority.
Shared understanding only matters if it changes what the system does. This pillar turns a shared world model into coordinated physical action — task allocation, macro-actions, multi-robot and robot-IoT coordination — and insists that the resulting decisions carry a trace a human can read and challenge.
The other four pillars are constrained by what a physical agent can actually compute, in time, on a power budget, in the field. This pillar develops compact models grounded in graph representations, distributes inference across the robot-IoT-edge continuum, and validates in simulation before hardware.
Interpretable Reasoning Framework for IoT–Robot Collaboration using Semantic Graphs
Japan Society for the Promotion of Science — KAKENHI, Early-Career Scientists · Principal Investigator · 2026 — 2028
Awarded
Domain-Adaptive Semantic Evidence for Connected Robotics
Tough Cyber-Physical AI Research Center — Research Project · Principal Investigator · 2026 — 2027
Awarded
Reasoning Over IoT and Robot Vision for Hazardous Scene Understanding
NVIDIA — Academic Grant Program (DGX Spark) · Principal Investigator · 2025
Awarded
Optimizing Robot Response: Using SLMs for Robotic Data Synthesis and LLMs for Interpretation
Tough Cyber-Physical AI Research Center — Joint Research Project · Principal Investigator · 2024 — 2025
Awarded
Knowledge-based Digital Twin Update using Multi-camera IoT with Federated Learning
Tough Cyber-Physical AI Research Center — Joint Research Project · Principal Investigator · 2023 — 2024
Awarded
Knowledge Acquisition for Autonomous Robots from Multimodal Information
Japan Science and Technology Agency — AIP Challenge · Principal Investigator · 2019 — 2020
Institutional context
This research is conducted within the Yamazaki–Bezerra Laboratory (Tohoku Robotic Systems Laboratory) at Tohoku University, led by Prof. Kimitoshi Yamazaki, where complementary research spans robotic manipulation, mobile manipulation, household robotics, Connected Robotics, and Physical AI. Connected Robotics and Edge AI is one of the laboratory’s four research pillars, and the direction my own research develops. Part of the work is carried out with the Tough Cyberphysical AI Research Center.