[Robot Hardware 01] - What Is Robot Hardware?
Robot hardware from a Physical AI perspective - Introduction
Robot Hardware from a Physical AI perspective
What Do We Mean by Robot Hardware?
When people hear “robot hardware,” they usually think first of visible parts: motors, frames, sensors, and machined components. In the dictionary sense, hardware simply means the physical pieces of a system.
But in a multi-jointed robot, hardware means much more. It includes the drive mechanisms, sensors, and control PC, but also the embedded systems that run high-frequency motor control, the network that connects the robot to its controller, physics-model-based control such as gravity compensation, and even floating-base CoM control or optimal-control-style stabilizers for legged robots. In robotics, all of these layers are often grouped together as the “hardware level.”
Strictly speaking, many of these pieces are software. But without them, the robot cannot even reach the point where experiments or data collection are possible.
So if we treat robot hardware as merely “a machine with many motors,” we miss the part that matters most. From the perspective of Physical AI, a robot is not just a machine that executes commands. It is an intelligent system that acts directly on the physical world. It applies force, the environment reacts, and the robot senses the result. Hardware sits at the center of that loop.
Even when two multi-jointed robots look similar, the way they interact with the world can be very different. With the same joint layout, changes in reducer type, material, and gear ratio alter backdrivability and mechanical impedance. Those properties decide whether the robot feels rigid or compliant at the moment of contact.
Once the robot’s mass distribution and inertia are added, the same control input can produce very different forces and accelerations. In tasks with repeated contact, such as manipulation and locomotion, those differences directly affect stability and success rate.
The same is true for sensing and control. Control frequency, communication delay, noise, sensor sensitivity, linearity, and hysteresis all shape the final behavior. Add structural stiffness, joint compliance, and friction, and a robot’s interaction with the world becomes the coupled result of many classical mechanical and electromagnetic properties.
AI and Robot Hardware
The complexity of robot hardware is not just a mechanical inconvenience. It is one of the fundamental barriers that modern AI methods run into. I think this shows up most clearly in two forms: the sim-to-real gap and the embodiment gap.
The first is the sim-to-real gap, a familiar problem in reinforcement learning for legged robots, drones, and dynamic control. Simulation cannot perfectly reproduce small but important physical effects such as friction, gearbox backlash, structural flexibility, communication delay, and vibration. If an AI model overfits to the clean physics of a simulator, even small real-world friction or oscillation can make the controller diverge or behave in unexpected ways. The forces a robot exchanges with the world are ultimately determined not by the simulator, but by the physical properties of the real hardware.
The second is the embodiment gap. This is a major obstacle for recent data-driven approaches that try to scale by collecting large datasets, such as VLA models and imitation learning.
Simple kinematic differences, such as link lengths or degrees of freedom, can often be handled with retargeting. The deeper problem is that every robot has its own dynamics and force relationships. Even if two robots follow the same trajectory, joint friction, actuator inertia, reducer backlash, and control bandwidth change the forces applied to the environment and the reactions that come back.
For example, data collected on a torque-controlled compliant robot does not transfer cleanly to a stiff position-controlled robot. The input-output relationship embedded in the data depends on the impedance of the hardware. When the platform changes, the physical rules of interaction change with it. That makes it difficult to reuse datasets or policies across robots, even when the task looks the same from the outside.
In practice, the limits of AI-based robot control often come less from the algorithm itself and more from the physical conditions imposed by the hardware. Without a deep understanding of hardware, it is hard to close the sim-to-real gap, and it is just as hard to build general datasets that survive the embodiment gap.
Understanding Robot Hardware
In this series, I want to break robot hardware down into layers. We will start from visible mechanical structures and actuators, move through drive mechanisms and sensors, and then work toward the low-level and physics-model-based control needed to make a multi-jointed robot move.
The goal is to understand what each layer does, where its limits come from, and how those limits affect real robot performance and AI-based control. I am organizing these notes from the perspective of someone who does not just use robots, but has to design, control, debug, and repair them.
Next post: [Robot Hardware 02] - Actuators (1): Motors
References
[1] https://www.ros.org/news/2016/10/hardware-robot-operating-system-h-ros.html