This proposal presents a strategic framework for strengthening research and development (R&D) of Physical AI, which emerges from the integration of AI and robotics. The objective is to advance robot technologies capable of adapting to real-world environments, thereby enhancing Japan's international competitiveness and societal relevance in the robotics domain. "Physical AI" refers to embodied artificial intelligence that acquires and develops intelligence through direct interaction with the physical environment via components such as sensors and actuators. We define a "Physical AI system" as a comprehensive system consisting of an AI robot that autonomously performs actions, learning, and decision-making, along with its supporting operational infrastructure. Unlike conventional cyber-centric AI systems, Physical AI systems must operate under physical and social constraints, enabling task execution and human-robot collaboration in the real world. In areas directly tied to people's lives--such as labor, caregiving, disaster response, and aging society support--embodied intelligence is essential for the practical and sustainable deployment of AI. Physical AI is expected to play a central role as the foundation for future robotics.

To systematically organize the capabilities and value required of Physical AI systems, this proposal defines three directions for future technological development from the perspective of AI robots:

• (1) enhancement of task execution,
• (2) adaptability to diverse environments, and
• (3) coexistence with humans.

Aligned with these directions, this proposal specifies the value characteristics that AI robots should exhibit and presents envisioned robotic models corresponding to each application domain:

• - Type P (Performance): AI robots capable of autonomously learning and accurately performing a wide variety of real-world tasks . These systems contribute to addressing labor shortages and skill transfer challenges by enabling precise industrial operations, replicating expert techniques, and expanding into fields such as healthcare, inspection, and services.
• - Type H (Humanoid): AI robots with the capacity to understand human intentions and actions, and to learn and adapt through interaction. These systems build natural collaborative environments and play vital roles in domains where human cooperation is essential, including caregiving, daily assistance, and education.
• - Type A (Adaptive): AI robots with robust capabilities to operate stably in harsh and dynamically changing environments. They are intended for deployment in outdoor tasks, agriculture, construction, disaster response, and infrastructure inspection, where human labor is difficult or dangerous.

Since the early 21st century, robotics has rapidly evolved through advancements in AI. In the 2010s, the adoption of machine learning and deep learning significantly improved perceptual capabilities such as image and speech recognition. In the 2020s, foundation models and generative AI have enabled the integrated processing of multimodal information--including language, vision, and motor actions--allowing robots to perform planning and generate behaviors. Leading software companies such as Google and OpenAI have begun applying multimodal foundation models to real-world robot control. Following this, hardware companies and public agencies in the U.S. and China have extended these models to bipedal and quadrupedal robots, initiating field demonstrations in factories and logistics. This has spurred cross-sector collaboration among researchers and engineers in software and hardware, addressing challenges such as environmental adaptability, real-time performance, cost, and safety.

Meanwhile, although Japan retains technological strength and market presence in industrial robotics, it lags behind the U.S. and China in applying generative AI to robotics. Japan's hardware excellence, grounded in real-world practice, now faces the test of maintaining its comparative advantage amidst rapidly advancing AI robotics abroad. Whether Japan can leverage the true strength of Physical AI─real-world adaptability─depends on timely and strategic response. Delay could prove critical.

Based on this context, the proposal identifies the following four core R&D challenges:

1. Development of adaptive Physical AI for real-world operation
Integrating AI and hardware to dynamically respond to tasks, environments, and robot-specific constraints, while overcoming challenges of real-time execution and energy efficiency at the edge.

2. Understanding embodied intelligence
Investigating the relationship between physical structure and intelligence, sensory feedback, and embodied cognition through constructive approaches rooted in cognitive developmental robotics.

3. Safety and trust in Physical AI systems
Establishing reliable systems through design incorporating risk assessment, fault tolerance, and recovery capabilities, along with safety guidelines and "by design" approaches.

4. Humanities and social science research on socio-economic impacts
Addressing social concerns such as safety perception, privacy, human-AI interaction, and responsibility allocation, and promoting interdisciplinary research and policy coordination based on the system's stage of development and degree of integration with society.

To effectively address these challenges, four complementary activities are essential:

• (1) promoting basic research,
• (2) evaluating socio-economic impacts,
• (3) building open research platforms, and
• (4) organizing competition-based testbeds simulating real-world conditions.

It is critical to establish a collaborative ecosystem supported by interdisciplinary research structures, standardized experimental environments, and evaluation frameworks reflecting realworld needs. Neutral support organizations, academic societies, research institutes, and industrial consortia must cooperate under institutional frameworks to sustain long-term R&D.

Physical AI R&D is expected to contribute to labor augmentation, workload reduction, and productivity enhancement across manufacturing, caregiving, education, transportation, environmental monitoring, and disaster response. Moreover, studies on embodied intelligence not only advance robotics, but also deepen our understanding of human cognition and serve as a foundation for academic development. Through promoting Physical AI and enhancing robots' adaptability to real-world environments, this proposal aims to establish a strategic foundation for strengthening Japan's global leadership in robotics. Particularly, deepening research on embodied intelligence may address limitations inherent in cyber-based AI and unlock new directions for technological innovation. These efforts are expected to accelerate the real-world implementation of robots and contribute to industrial competitiveness and societal well-being.

• (in Japanese) Physical AI System - Integration of Embodied AI and Robotics -