SAIL: Test-Time Scaling for In-Context Imitation Learning with VLM
IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2026
accepted
Makoto Sato 佐藤 誠人
Ph.D. Student · Matsuo-Iwasawa Lab, The University of Tokyo
I am a second-year Ph.D. student in the Department of Technology Management for Innovation, Graduate School of Engineering, The University of Tokyo. My research focuses on robot learning, with an emphasis on scalable imitation learning and test-time scaling for robot foundation models. Previously, I worked on task and motion planning for contact-rich manipulation at NAIST, and on multimodal object identification at Saitama University. I have also collaborated with Sakana AI, AIRoA, Matsuo Institute, and AIST.
Robot learningImitation learningTAMPWorld modelsFoundation modelsContact-rich manipulation
News
- Started a research internship at NVIDIA.
- SAIL — test-time scaling for in-context imitation learning with VLMs — accepted at the 2026 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS).
- Joined Sakana AI as a research intern.
- Joined AI Robot Association (AIRoA) as a research collaborator.
- Started Ph.D. at the Matsuo-Iwasawa Lab, The University of Tokyo.
- Paper accepted at SCI 2024 — TAMP with residual reinforcement learning for long-horizon manipulation.
Publications full list →
Task and Motion Planning Using Residual Reinforcement Learning for Long-Horizon Precise Object Manipulation
SCI, 2024
figure
Construction and Validation of Action-Conditioned VideoGPT
JSAI, 2023
Scaling Laws of Dataset Size for VideoGPT
JSAI, 2023
Scaling Laws of Model Size for World Models
JSAI, 2023
Imitation Learning with Mid-Level Representations for Object Rearrangement
JSAI, 2022
Projects research themes →
video
Exploratory Task and Motion Planning Using Uncertainty-Aware VLM
Studying how uncertainty-aware vision-language models can generate exploratory subgoals for task and motion planning.
Large-Scale Demonstration Generation for VLA via Generative Simulation
Building data generation pipelines for training autonomous robot foundation models using generative simulation.
Large-Scale Imitation Learning for Bimanual Dexterous Manipulation
Exploring scalable imitation learning for humanoid and dexterous manipulation in simulation-rich environments.
Long-Horizon Manipulation with Task and Motion Planning
Investigating planning-driven robot manipulation for long-horizon tasks that require precise interaction and sequencing.
Quadruped Locomotion for Unitree A1 via Massively Parallel PPO
Training the Unitree A1 quadruped to walk by running PPO across thousands of parallel simulation environments for fast, scalable reinforcement learning of locomotion.
Multimodal Object Identification with Vision and Piezoelectric Vibration Sensing
Undergraduate research on identifying objects from images and piezo-film (vibration) sensor signals, fusing the two modalities with a Siamese network for robust recognition.
Bio
Education
-
The University of Tokyo Ph.D., Technology Management for Innovation
-
NAIST M.Eng., Information Science
-
Saitama University B.Eng., Mechanical Engineering and System Design
Experience
-
NVIDIAResearch intern
-
Sakana AI, Inc.Research intern
-
AI Robot Association (AIRoA)Research collaborator
-
Matsuo Institute, Inc.Research engineer
-
Sony Semiconductor SolutionsFull-time engineer
-
AISTResearch group member
Teaching
-
Physical AI Spring Seminar Kinematics and AI models
-
Deep RL Spring Seminar Model-based RL and world models
-
World Model Course Simulation and computer graphics
Toolbox
Python · C++ · PyTorch · ROS · Docker · MuJoCo · Isaac Sim · Newton · PyBullet · SAPIEN · Gazebo · CARLA
Contact
makoto.sato@weblab.t.u-tokyo.ac.jp · github.com/makolon · Tokyo, Japan