We aim to develop a versatile humanoid robot, AIREC (AI‑driven Robot for Embrace and Care), capable of supporting humans in real‑world applications such as welfare and caregiving, where severe labor shortages are imminent. AIREC will perform a wide range of highly complex physical tasks and engage in real‑world communication with humans through emotions generated from its own embodied system. Specifically, we will pursue innovation of high‑power humanoid robot bodies designed for integration with end‑to‑end AI, and realization of diverse physical tasks through deep predictive learning and robot foundation models, thereby advancing AI and robotics together. Through these efforts, we will implement physical support tasks that involve interaction with humans.

We therefore aim to develop general-purpose autonomous humanoids that grow progressively through multi-layered instruction — hands-on guidance, learning by imitation, and verbal instruction — and through multi-layered inheritance of experience and skills from humans to robots, from robots to other robots, and from robots to humans. We will evaluate their adaptability, versatility, and ability to pass on what they have learned. Specifically, we pursue R&D on humanoids suited to instruction and inheritance, foundational systems that enable instruction and inheritance, task execution and real-world deployment based on instruction and inheritance, and next-generation community building. By 2050, we aim to realize humanoids that grow together with people and other robots by teaching and learning from one another, and that can take on all physical tasks that could reasonably be entrusted to part-time workers.

This research aims to realize a humanoid robot capable of performing real physical work in unstructured and challenging environments such as construction sites. Its novelty lies in the integrated development of: (i) lightweight, high-power, and highly durable hardware; (ii) locomotion and whole-body motion planning that fuses model-based control with learning; (iii) manipulation skill acquisition that structures and extends the tacit knowledge of skilled workers; and (iv) human–humanoid and multi-robot collaborative intelligence enabled by multimodal AI. Using actual construction and manufacturing sites as demonstration fields, we will verify the system in a stepwise manner. We will establish the core technological foundation required for practical deployment at limited work sites by 2030 and for the realization of general-purpose autonomous humanoid AI robots by 2050. This project will make a significant contribution to resolving the increasingly severe labor shortage and to achieving safe and sustainable on-site operations.

This project aims to realize general-purpose robots that perform diverse non-routine tasks at human-level speed, addressing severe labor shortages. We fuse the recognition and decision-making capabilities of Vision-Language-Action (VLA) models with advanced force-control technologies based on bilateral control, so that robots adapt flexibly to on-site tasks through interactive feedback. We also develop bilateral control systems with a 15 Hz control bandwidth from the motor driver and gear levels, building a robot platform that is taught and adapts on site.

This project aims to develop robotic intelligence that understands object properties and possible actions from forces, motions, and energy exchanged through physical interaction. By integrating contact, friction, and deformation information across intelligence, embodiment, and simulation, we seek to build general-purpose autonomous robots that can safely complete tasks in unknown or disturbed environments and adapt flexibly to changes in people and surroundings.

To achieve AI-driven robotic automation in agriculture, key technical components—biomimetics, multimodal integration centered on visuo-tactile sensing, and meta-learning—play an indispensable role in realizing general-purpose embodied intelligence. We deploy "Artemy 2.0," created by adding tactile sensing capabilities to DENSO's agricultural robot Artemy, for early data collection. This drives the evolution into "Artemy 3.0," featuring enhanced embodied intelligence and wrist dexterity capable of executing harvesting, leaf pruning, and crop training. Furthermore, through "Agricultural Physical Foundation Model", we will realize multimodal integrated agricultural humanoids by combining rapid adaptation to unknown crops with physical adaptation driven by metacognition. By 2050, we aim to transform agriculture from "strenuous labor" into an "advanced intelligent industry" where humans and robots creatively coexist.

We aim to establish a framework that enables humanoid robots to autonomously perceive their surroundings, explore appropriate actions, and learn through accumulated experience. Instead of relying on humans to explicitly teach individual movements, the robot will determine what to try next and progressively expand its behavioral capabilities through its own interaction with the environment. To achieve this, we will integrate a high-level mechanism for understanding the environment and selecting appropriate actions with a low-level mechanism for generating and executing whole-body movements according to the current situation. By tightly coupling these processes, we seek to build a foundation for humanoid robots that can adapt flexibly to diverse environments, continuously acquire new capabilities, and autonomously improve through experience.

This project aims to achieve a fundamental advancement in general-purpose autonomous humanoid AI robots by establishing core technologies centered on whole-body tactile sensing using electrical impedance tomography (EIT) and overcoming limitations in conventional robotic body structures. Next-generation EIT-based tactile sensors offering wide-area coverage, high spatial density, and high sensitivity will be integrated across the entire body of a humanoid AI robot using flexible functional materials. By processing vast amounts of tactile data in real time through attention mechanisms, the project will develop a new model of embodied intelligence that enables the robot to sense, act, understand, and adapt using its entire body.

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.

Our project aims to realize and electronic skin (e-skin) that mimics tactile perception and cognitive functions of human fingertips, and to develop a robotic hand equipped with this e-skin. Unlike conventional rigid sensors, e-skin is a soft and flexible sensor sheet that can bend and deform like human skin. In this project, we will develop and integrate the e-skin based on previously developed by the PM (an example is shown in the lower right figure) with machine learning to reproduce the ambiguous and subjective aspects of human tactile perception, such as slip, roughness, softness, and tactile feel. Furthermore, by integrating this e-skin system into a robotic hand, we aim to realize a system that can learn and adapt, as humans do, while recognizing the state of the object being grasped. Through this research, we aim to establish fundamental technologies for a future society in which humans and robots can live together naturally, with humanoid robotics possessing human-line tactile perception and sensitivity and becoming part of the family.

This research focuses on the hands and arms—the critical points of physical contact that enable AI-powered humanoid robots to work safely in people’s everyday living spaces. Our goal is to develop hands and arms that make grasping failures less likely to lead to serious accidents. Going beyond reliance on sensors and stop functions alone, we embed mechanisms inspired by the human body into the robot’s core physical structure. These include underactuation, which allows multiple joints to move with fewer actuators, and mechanical coupling that coordinates their motion. By building safety into the mechanism itself, we aim to create limbs that are less likely to exert dangerous forces, even when control systems malfunction or mistakes occur during learning.

This project aims to develop a humanoid body design with wheeled feet that can move stably across different speed ranges and use cases, from pedestrian areas such as sidewalks to roadside and delivery environments. In addition to energy-efficient wheeled locomotion, the robot will use the degrees of freedom of its upper body and legs for posture stabilization, establishing a technological foundation for safe and smooth mobility in public spaces shared with people. We will also evaluate the robot's behavior when passing or accompanying people and feed the results into design, control, and ELSI, with the aim of establishing a new locomotion principle that spans the "last mile" to the "last step."

Toward the realization of robots that continuously learn from the environment in which they are embedded, this project explores the principle of “learning to learn”: learning to obtain information relevant to the situation by spontaneously acting on the environment. We focus on tasks in which a robot uses tools on the spot to organize purposive encounters with the environment. Drawing on the approaches of reservoir computing and embodied intelligence, we aim to open a path toward robots that actively move their bodies to tune their attention to the information that specifies affordances—the possibilities for action offered by the human environment.

This study aims to establish fundamental technologies for equipping robots with the ability to “infer” how to provide movement assistance without causing discomfort to the user. Humans have the ability to intuitively “infer” the movement speed and level of force that are comfortable for others. By estimating physiological changes associated with discomfort from biosignals, we hypothesize that humanoid robots can be equipped with the ability to “infer” natural and comfortable assistive movements. To achieve this goal, we will identify and validate physiologically grounded biomarkers of discomfort through multimodal biosignal measurements in humans and non-human primates, and develop practical technologies* for estimating natural assistive movements. (*patent pending)