Although robotic automation is advancing quickly, most systems have difficulty adjusting their pre-trained movements to environments with objects that vary in stiffness or weight. To address this, a team of researchers in Japan has created an adaptive motion reproduction system based on Gaussian process regression.
Their approach models the relationship between human motion and object characteristics, allowing robots to accurately mimic human grasping actions. It can achieve this with minimal training data and handle unfamiliar objects with impressive precision and efficiency.
Obstacles to Robotic Flexibility
“Rapid advancements in robotic automation have the potential to transform industries and enhance our lives by taking over tasks that are dangerous, physically strenuous, or monotonous for humans.”
Although current robots perform exceptionally well in structured settings like assembly lines, the true challenge of automation is operating in unpredictable, dynamic environments, such as cooking, elderly care, or exploration.
Achieving this requires overcoming a major obstacle: enabling robots to sense and adapt through touch. Unlike human hands, which naturally adjust their grip to objects of varying weight, texture, or stiffness, most robotic systems still lack this essential adaptability.
Progress in Motion Replication Technologies
To equip machines with advanced human-like dexterity, researchers have created a variety of motion reproduction systems (MRSs). These systems focus on precisely capturing human movements and replicating them in robots through teleoperation.
Nevertheless, MRSs often struggle when the characteristics of the object being manipulated differ from those used during the original motion recording. This reduces the flexibility of MRSs and, consequently, limits the broader usability of robots.
To tackle this core challenge, a research team in Japan has created an innovative system capable of adaptively modeling and replicating intricate human movements.
The study was spearheaded by Master’s student Akira Takakura from the Graduate School of Science and Technology at Keio University, with contributions from Associate Professor Takahiro Nozaki of the Department of System Design Engineering, Doctoral student Kazuki Yane, Professor Emeritus Shuichi Adachi of Keio University, and Assistant Professor Tomoya Kitamura from Tokyo University of Science, Japan.
Enhancing Adaptability Through Gaussian Process Regression
The team’s key innovation was moving beyond linear modeling approaches and adopting Gaussian process regression (GPR), a method capable of capturing complex nonlinear relationships even from limited training data.
By recording human grasping motions across a variety of objects, the GPR model learned how an object’s “environmental stiffness” relates to the position and force commands applied by humans. This process effectively uncovers the underlying human motion intent, or “human stiffness,” enabling the robot to perform suitable actions with objects it has never handled before.
“Equipping robots with the ability to manipulate everyday objects is crucial for allowing them to interact naturally with their environment and respond appropriately to encountered forces,” says Dr. Nozaki.
System Testing and Performance Results
To validate their method, the researchers compared it with traditional MRSs, linear interpolation, and a standard imitation learning model.
The proposed GPR-based system showed substantially improved performance in generating accurate motion commands for both interpolation and extrapolation tasks.
For interpolation—predicting motions for objects with stiffness values within the training range—the method reduced the average root-mean-square error (RMSE) by at least 40% for position and 34% for force.
For extrapolation—handling objects stiffer or softer than those in the training set—the approach remained highly effective, achieving a 74% reduction in position RMSE. Overall, the GPR-based method significantly outperformed all other tested approaches.
Applications in Industry and the Evolution of Robotics
By effectively modeling human–object interactions using minimal training data, this novel approach to MRSs enables the generation of precise and dexterous motion commands for a diverse array of objects. Its capacity to capture and replicate complex human skills allows robots to operate beyond rigid, pre-defined scenarios, paving the way for more sophisticated and adaptable services.
“Because this technology requires only a small dataset and reduces the costs associated with machine learning, it has broad potential across many industries,” explains Mr. Takakura. “For example, life-support robots, which must adjust their movements to different targets each time, could greatly benefit, and companies that previously struggled to implement machine learning due to large data requirements may now find it more accessible.”
Read the original article on: Tech Xplore
Read more: China Unveils a Humanoid Robot with Smooth, Human-like Balance
