Summary of the project

We will develop technologies for generating cooperative motions that enable robots working with humans to continue operating safely despite changes in the environment and unexpected contact, and to recover from disturbances. We will integrate Koopman analysis, a mathematical approach to analyzing nonlinear dynamics, with dynamic movement primitives (DMPs), a conventional framework for stable motion generation, and establish a method for learning reusable motion representations from diverse data. This will enable robots to adapt rapidly to changing situations during task execution by combining motions and updating intermediate goals, without requiring large-scale retraining. We will also develop technologies for reducing model size while preserving stability, thereby enabling low-latency, low-power, and highly reliable edge implementation.

Milestone by year 2028

By 2028, we will integrate a data-driven method for generating stable motions with prediction and control algorithms that select, compose, and switch intermediate goals and motions according to the situation. Using a robotic arm to perform a sequence of cooperative tasks, we will experimentally compare the developed method with conventional ones based on classical DMPs, imitation learning, and reinforcement learning, and demonstrate that it can complete tasks safely without large-scale retraining. We will also execute a lightweight model in closed loop in an edge-equivalent environment, demonstrate the feasibility of satisfying control-cycle timing requirements while maintaining coordination performance, and calibrate the simulation conditions using real-robot data to clarify the applicable range and limitations.

R&D Items

• [1] Establishment of theoretical foundations for operator-theoretic, data-driven dynamic movement primitives for versatile and stable motion generation
• [2] Development of integrated prediction and control algorithms for rapid, adaptive motion generation in response to changing conditions
• [3] Development of distillation and transfer learning algorithms for motion representation bases to enable context-specific edge deployment
• [4] Experimental validation of the developed methods using both testbeds representative of edge deployment and robotic arms

Performers

PDF Download

• Summary of the project (450 KB)