Research

Project

• 
  Inclusive robotic foundation model (JST CRONOS)
  We develop a world model connected to a foundation model that can optimize the actions for various robots in response to language instrutions.
  • Mapping between latent action space common among robots and robot-specific action spaces
  • Lightweight hypernetworks that switch internal state representation according to language instrutions
  • Learning world model interpreted as multi-objective optimization

Reinforcement learning

• 
  Reinforcement learning as probabilistic inference
  We derive and analyze various values that can be found by interpreting RL as a kind of probabilistic inference problem.
  • Optimization of discount factor according to events
  • Weber-Fechner law in TD learning
  • Distributional model with regular optimism and pessimism
  • Theoretically-grounded optimistic learning
  • Integration with feedback error learning
• 
  Model-based reinforcement learning
  We study learning theory of world model and its application to model predictive control in real time.
  • Integration of model predictive control, imitation learning, and reinforcement learning
  • Adversarial learning with moderate robustness and less conservativeness
  • Model predictive control with efficient convergence to a suboptimal solution
  • Extraction of sparse low-dimensional latent space
• 
  Stabilization of deep reinforcement learning
  We develop stabilization techniques that enable deep reinforcement learning to optimize policies stably.
  • Optimization of divergence direction in policy improvement
  • Analysis of experience replayable conditions
  • Stable and fast target network updates
  • Regularization of local Lipschitz continuity of policy and value functions

Imitation learning

• 
  Utilization of imperfect demonstration
  We develop imitation learning methodologies in the absence of sufficient quality, modality, and quantity of demonstration data.
  • Guided reinforcement learning with rough target trajectories as constraints
  • Spatiotemporal partial imitation through self-paced learning
  • Robust behavioral cloning for outliers based on Tsallis statistics
  • Safe and efficient framework to compensate for missing action data
• 
  Interactive imitation learning
  We study a framework in which the demonstrator and agent collaborate to collect data to be imitated.
  • Lifelong learning of non-stationary multiple tasks
  • Active exploration without sacrificing the sense of agency
  • Optimal consensus decision-making among multiple actions with confidences

Other machine leanring

• 
  Time-series data processing
  We design models to process time-series data, which is important for analyzing and representing robot and human motions.
  • Learning theory of recurrent neural networks based on variational inference
  • Design of reservoir computing in complex domain that embeds the representation of periodicity
  • Approximated implementation of neuron dynamics following a power law
• 
  Stochastic gradient descent
  We improve the performance of the stochastic gradient descent, which is a core technique in deep learning.
  • Novel interpretation of Lookahead optimizer
  • Robustness to noise and outliers based on Student's t-distribution
  • Improvement of AMSGrad to make it applicable to non-stationary problems
• 
  Lifelong (continual) learning
  We develop fundamental technologies for autonomous robots to keep learning continuously.
  • Stochastic gradient descent without catastrophic forgetting
  • Generalization and hierarchization of reservoir sampling
  • Optimization of balancing rehearsal and function regularization
  • Fractal model for continual learning of multiple tasks

Domain-oriented

• 
  Humanoid robots
  We study motion control of humanoid robots, which are expected to play an active role in human society as general-purpose robots.
  • Integration of whole-body model predictive control and machine learning
  • Dancing with a partner using whole-body skin sensors
  • Unified controller for bipedal walking and running and continuous transition between them
• 
  Human assistive robot
  We develop AI technologies to assist various people.
  • Extracting and understanding the tacit knowledge hidden in subjective evaluation
  • Surgical assistance with learning-based model predictive control
  • Learning assistance in embryo manipulation with extracted expert skills
  • Locomotion assistance based on recognition of human movement behaviors