Generative AI such as ChatGPT excels at creating text and images on a screen. In contrast, AI that has a body, such as a robot, and grasps objects or moves around in the real world is called Physical AI (embodied AI). This article explains the basics of Physical AI and how to start learning it.
What is Physical AI?
Physical AI is AI that treats the whole loop as one: sensing the surrounding environment, deciding the next action with AI, and actually moving with motors and arms. It is also called "embodied AI."
While generative AI stays within the world of words and data, Physical AI must deliver results in the real world, governed by gravity, friction and contact. That is why robot hardware, control and data collected in real environments matter as much as the AI model itself.
The three building blocks of Physical AI
- Perception: understanding the surroundings and objects with cameras, LiDAR, IMUs and tactile sensors
- Decision: understanding instructions and planning tasks with large language models (LLMs) and vision models
- Action: carrying out the work with mobile bases, robot arms and robot hands
Today, in research and education settings, you can say "move the red box to the right shelf" and an LLM breaks the task down while the robot combines image recognition and motion planning to execute it.
Why Physical AI is attracting attention now
There are three main reasons. First, advances in LLMs let robots understand human language and plan tasks flexibly. Second, human-like hardware such as humanoids and dexterous robot hands has become affordable. Third, imitation learning, which trains robots on recorded human demonstrations, is becoming an established method.
How to learn Physical AI
Physical AI cannot be mastered through software alone. Hands-on experience reading sensor data, driving robots with ROS and combining them with AI models is essential. Universities, technical colleges and corporate training programs typically follow these steps:
- STEP 1: learn mobile robot fundamentals such as ROS, SLAM and PID control
- STEP 2: understand robot structure, actuation and control module by module
- STEP 3: combine LLMs and image recognition with a robot arm to integrate perception through action
- STEP 4: develop a robot system as a team through project-based learning
Examples of educational platforms for each step:
- UNI-WR2S: desktop robot for learning ROS and SLAM (STEP 1)
- GX-MAT-09S: a modular embodied AI robot design kit (STEP 2)
- RAI-M4: Physical AI learning platform with LLM integration (STEP 3)
- RAI-P4: task-planning training platform with a robot arm (STEP 3)
- PBL training system: project-based development training (STEP 4)
Applications in research and industry
In research and industry, Physical AI is starting to be used to automate dexterous manual work and to develop service robots and humanoids. For example, a data glove records human hand motion and touch while a multi-fingered robot hand reproduces it and training data is collected.
- Wuji Glove: data glove that records hand motion and touch
- Wuji Hand 2: 20-DoF dexterous robot hand
- UX052: mobile base for wheeled humanoids
- BXI Elf3: humanoid body ODM
Summary
Physical AI performs perception, decision-making and action as one loop in the real world. Thanks to advances in generative AI and more accessible robot hardware, it is now within reach for education, research and industry alike. HumansX provides robot platforms for learning, building and deploying Physical AI. Feel free to contact us about adoption or course design.