This image depicts the real‑time transfer of a human’s motion to a robotic avatar, enabling the latter to perform a dexterous task. Credit: Keio University Global Research Institute (KGRI)

Despite rapid robotic automation advancements, most systems struggle to adapt their pre‑trained movements to dynamic environments with objects of varying stiffness or weight. To tackle this challenge, researchers from Japan have developed an adaptive motion reproduction system using Gaussian process regression.

By learning the relationship between human motion and object properties, their method enables robots to accurately replicate human grasping behaviors using small training datasets and manipulate unfamiliar objects with remarkable precision and efficiency.

Challenges in robotic adaptability

Accelerating progress in robotic automation promises to revolutionize industries and improve our lives by replacing humans in risky, physically demanding, or repetitive tasks.

While existing robots already excel in controlled environments such as assembly lines, the ultimate frontier of automation lies in dynamic environments found in tasks such as cooking, assisting the elderly, and exploration.

To realize this goal, one of the key barriers is making robots capable of adapting to touch. Unlike human hands, which intuitively adjust their grip for objects of unknown weight, friction, or stiffness, most robotic systems lack this crucial form of adaptability.

Advances in motion reproduction systems

To transfer sophisticated human dexterity to machines, researchers have developed various motion reproduction systems (MRSs). These are centered around accurately recording human movements and recreating them in robots via teleoperation.

However, MRSs tend to encounter problems if the properties of the object being handled change or do not match those of the recorded movement. This limits the versatility of MRSs and, in turn, the applicability of robots in general.

To address this fundamental challenge, a research team from Japan has developed a novel system designed to adaptively model and reproduce complex human motions.

The study was led by Master’s student Akira Takakura (Graduate School of Science and Technology, Keio University) and co‑authored by Associate Professor Takahiro Nozaki (Department of System Design Engineering, Keio University), Doctoral student Kazuki Yane, Professor Emeritus Shuichi Adachi (Keio University), and Assistant Professor Tomoya Kitamura (Tokyo University of Science, Japan).

Their paper is published in IEEE Transactions on Industrial Electronics.

How Gaussian process regression improves adaptability

The team’s core breakthrough was moving past linear modeling strategies and instead using Gaussian process regression (GPR). This regression technique can accurately map complex nonlinear relationships, even with a small amount of training data.

By recording human grasping motions for multiple objects, the GPR model was trained to identify the relationship between the object’s “environmental stiffness” and the necessary position and force commands issued by the human. In turn, this process effectively reveals the human’s underlying motion intention, or “human stiffness”, allowing the robot to generate appropriate motion for objects it has never encountered.

“Developing the ability to manipulate commonplace objects in robots is essential for enabling them to interact with objects in daily life and respond appropriately to the forces they encounter,” explains Dr. Nozaki.

Testing and results of the new system

To validate their approach, the researchers tested it against conventional MRSs, linear interpolation, and a typical imitation‑learning model.

The proposed GPR system demonstrated significantly enhanced performance in reproducing accurate motion commands for both interpolation and extrapolation.

• Interpolation (handling objects with stiffness within the training set limits) – average root‑mean‑square error (RMSE) reduced by at least 40 % for position and 34 % for force.

• Extrapolation (objects harder or softer than those in the training set) – 74 % reduction in position RMSE.

Overall, the GPR‑based method markedly outperformed all other methods.

Implications for industry and future robotics

By accurately modeling human–object interactions with minimal training data, this new take on MRSs will help generate dexterous motion commands for a wide range of objects. This ability to capture and recreate complex human skills will ultimately enable robots to move beyond rigid contexts and toward providing more sophisticated services.

“Since this technology works with a small amount of data and lowers the cost of machine learning, it has potential applications across a wide range of industries, including life‑support robots, which must adapt their movements to different targets each time, and it can lower the bar for companies that have been unable to adopt machine learning due to the need for large amounts of training data,” says Mr. Takakura.

More information:

Akira Takakura et al., “Motion Reproduction System for Environmental Impedance Variation via Data‑Driven Identification of Human Stiffness,” IEEE Transactions on Industrial Electronics (2025). DOI: 10.1109/tie.2025.3626633

Provided by Keio University Global Research Institute (KGRI)

Citation: Adaptive motion system helps robots achieve human‑like dexterity with minimal data (2026, January 13) retrieved 13 January 2026 from https://techxplore.com/news/2026-01-motion-robots-human-dexterity-minimal.html

This document is subject to copyright. Apart from any fair dealing for the purpose of private study or research, no part may be reproduced without written permission. The content is provided for information purposes only.