M.S. student in the Hagiwara Lab at Soka University · Robotics researcher with Team SOBITS · Full-stack AI engineer at dot.lab
I build embodied-AI systems that connect language-level reasoning to uncertain evidence in the real world. My work spans service-robot planning, ROS 2 integration, perception and control, plus production AI software.
- Robot planning: long-horizon task planning from natural language, including when a robot should act, observe, or replan.
- Real-robot engineering: ROS 2 systems that connect navigation, manipulation, perception, and human interaction.
- Production AI: reliable client-facing tools, automation, APIs, access control, and evaluation.
| Project | Contribution | Evidence |
|---|---|---|
| LLM chaining for service robots | Designed and evaluated a two-stage planner for RoboCup@Home GPSR tasks; validated on a Toyota HSR. | RSJ 2026 presentation · GCCE 2026 — accepted, oral presentation |
| RoboCup service-robot systems | Integrated planning, perception, navigation, manipulation, and interaction with Team SOBITS; contributed targeted ROS 2 reliability improvements. | Team SOBITS · Soka University RoboCup 2026 report |
| Evidence-grounded planning | Classifying environmental evidence as supported, absent, unknown, stale, or ambiguous so the robot can act, observe, or replan. In progress; real-robot demo targeted Nov 2026. | Hagiwara Lab, HSR |
| SOBIT-X — VLA adaptation | Leading the HSR adaptation direction on AIRoA's π0.5 HSR base: lab-robot interface, task-data alignment, fine-tuning, real-robot evaluation. | Team SOBITS, contributor |
| MCP robot-skill interface | Wrapping existing ROS 2 robot skills as MCP servers so agent planners can invoke navigation, perception, and manipulation through one interface. In progress. | Hagiwara Lab |
| Production AI at dot.lab | Build full-stack AI tools, from user flows and APIs to evaluation, billing, persistence, and recovery. | Client work is private and described in my portfolio. |
| Open-source ROS 2 packages (Team SOBITS) | yamnet_ros — 521-class real-time sound-event detection · sam3_ros — text-prompted SAM3 segmentation · sobits_follower — LiDAR + RGB-D person following · sobit_pro — C++ joint and pose control | Public ROS 2 packages |
| Smart Meeting Audio Assistant | Integration owner: designed the typed SystemSnapshot contract and mock-data layer so teammates could build against a stable interface, then built both frontends, audio capture, Whisper transcription, and visualization. |
Repo — MIT, Python |
| AI Robot Book — Chapter 3 | Contributing to the English edition on speech recognition and synthesis, with ROS 2 action-based sample programs. | Chapter 3 repository |
Robotics: ROS 2 · Nav2 · TF2 · SMACH · URDF/xacro · PCL · YOLO/SAM3 integration · LiDAR/RGB-D tracking
AI and software: LLM systems · evaluation · MCP · Whisper · DeepEval · FastAPI · React/Next.js · Hono/Bun · PostgreSQL/Supabase · Playwright
Training & models: PyTorch (YOLO detection weights) · TensorFlow (YAMNet integration)
- Presented "Design of a Stepwise Instruction-Processing Structure for GPSR Tasks Using LLM Chaining" at the 13th Intelligent Home Robotics Workshop, University of Tokyo, December 2025.
- First author: Evaluation of the Effectiveness of LLM Chaining for Long-Horizon Task Planning in General-Purpose Service Robots — RSJ 2026.
- First author: Design and Evaluation of LLM Chaining-Based Task Planning for General Purpose Service Robots — IEEE GCCE 2026 — accepted, oral presentation.
- Featured in CAPA magazine's Portuguese-language interview on AI and robotics for Brazilians in Japan.
Portfolio · LinkedIn · Team SOBITS · Email