Research

Research in scalable robot learning

I work on robot learning systems that benefit from richer demonstrations, stronger simulation, and more effective inference-time decision making. My current Ph.D. research at The University of Tokyo focuses on scalable imitation learning and test-time scaling for robot foundation models.

See Publications CV and Biography Contact

Current Focus

  • Scalable imitation learning Learning from large and diverse robot demonstrations.
  • Test-time scaling Using additional inference-time compute for embodied decision making.
  • Long-horizon manipulation Combining planning and learning for contact-rich tasks.
Themes

Core research directions

These themes connect my work across robot learning, embodied intelligence, and scalable machine learning systems.

Scalable Imitation Learning

Building manipulation policies from large-scale demonstrations and representation-rich observations that generalize across tasks and embodiments.

Test-Time Scaling

Studying how extra inference-time computation can improve in-context reasoning, action selection, and recovery behavior for robot models.

Long-Horizon Manipulation

Combining task and motion planning with learning-based control for precise manipulation that requires sequencing, memory, and recovery.

Simulation and World Models

Using generative simulation, embodied datasets, and world models to scale training signals for more general robot policies.

Projects

Recent project threads

My recent work spans simulation-led data generation, uncertainty-aware planning, and policy learning for dexterous and long-horizon tasks.

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 that make autonomous robot foundation models easier to train at scale.

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 manipulation for long-horizon tasks that require precise interaction and sequencing.

Research ingredients

  • Robot learning
  • Imitation learning
  • Task and motion planning
  • Reinforcement learning
  • World models
  • Robot foundation models
  • Vision-language models
  • Physical AI

Collaboration interests

  • Academic collaboration Embodied intelligence, scalable robot learning, and evaluation.
  • Industry projects Applied physical AI systems, training infrastructure, and benchmarking for manipulation.
  • Seminars and discussions Talks on robot learning, world models, and inference-time scaling.