Research Topics

We work on neuro-musculo-skeletal modeling and reinforcement learning for human motion control and interactive robotics.

Related areas

AI for Health, Machine Learning, Neural Engineering, Robotics

MS-Emulator

Musculoskeletal Control Motion Imitation World Models

A general whole-body controller that coordinates 700 muscle actuators in MS-Human-700 with one shared policy across diverse human motions. Learning through differentiable world-model imagination, value-guided flow exploration, and adversarial rewards enables efficient, accurate motion tracking.

MS-Human-700

Musculoskeletal Model Neural Control Biomechanics

A whole-body human musculoskeletal model with 700 muscle-tendon units, built to capture realistic human movement at full-body scale. It serves as a biomechanically grounded platform for studying motor control and learning controllers for human-like motion.

FlyLearn

Connectomics Embodied Learning Neural Model

FlyLearn explores whole-brain connectomics for embodied learning. Using the Drosophila connectome as a foundation, we develop neural architectures and learning algorithms for efficient information processing and embodied control.

Safe Optimization

Bayesian Optimization Safe Learning

Efficient black-box optimization under safety constraints, where unsafe evaluations must be avoided throughout the search. The methods balance exploration and constraint satisfaction so that performance improves without violating safety requirements.

Preference-based Optimization

Bayesian Optimization Preference Learning

Optimization driven by preference feedback rather than explicit numerical objectives, where users compare candidates instead of scoring them. This makes it well suited to tuning systems whose quality is hard to quantify but easy to judge.

High-dimensional Optimization

Bayesian Optimization High-dimensional

Scalable optimization methods for high-dimensional problems, where classical Bayesian optimization struggles. The work develops techniques that remain sample-efficient as the search space grows.

Preference-based Learning

Reinforcement Learning Exoskeleton Human-in-the-loop

A personalized gait optimization framework for lower-body exoskeletons that learns from human feedback. It adapts assistance to each user's preferences, improving comfort and walking performance during interaction.

Wheeled-legged Quadruped

Reinforcement Learning Legged Robotics Locomotion

Learning adaptive locomotion for wheeled-legged quadrupeds that combine wheels and legs for versatile movement. The controllers switch between rolling and stepping to traverse varied terrain efficiently.