Most of us think of robots as machines built to perform exactly one job over and over, like welding a car part or moving a box in a warehouse. If that robot moves an inch out of place, it fails because it lacks the basic ability to adapt. Companies like X Square Robot are trying to change this by building what they call a foundation stack—a set of shared digital tools that could eventually give robots the flexibility to handle messy, unpredictable environments like our living rooms.
The team at X Square Robot is moving away from the traditional way of building robots, where different experts program separate parts of the machine for vision, planning, and movement. Instead, they are using a unified architecture that learns how to see, reason, and act all at once. Crucially, they are releasing parts of this system as open-source, meaning other researchers can download and build upon their work. They have also developed a new way to collect training data. Rather than using expensive equipment to drive a robot remotely, they use humans wearing motion-tracking rigs to perform tasks. This allows the robot to learn from the fluidity of natural human movement.
Rethinking how robots learn
To teach a robot, you need data. In the past, engineers relied on teleoperation, where a person plays a video game with a robot arm, guiding it through every single movement. This is slow and difficult. X Square Robot does something different: they use a wearable rig that records human hand movements. After the person does the task, the system checks if the robot can actually perform those same motions successfully. If the robot misses, the data is discarded. This process creates a cleaner, higher-quality dataset than simply recording every movement a robot makes.
Inside the AI, the machine doesn't just see a video feed; it translates movements into semantic events, such as reaching, grasping, or placing. Think of this like how you might describe a recipe: you don't track every millisecond of a chef’s hand motion; you track the goal of each step. By organizing learning around these human-understandable goals rather than just clocking time, the robot is better at correcting itself if it makes a mistake midway through a task. This mimics how humans handle life—we don't perform a sequence of rigid frames; we perform a sequence of intentions.
The biggest hurdle to having a robot in your home isn't getting it to load a dishwasher once—it is reliability. If a robot drops a glass or gets stuck because a chair was moved, a normal robot gives up. A truly capable system needs to know how to recover from failure because a home is not a controlled factory floor. By focusing on how robots can safely and independently fix their own mistakes, researchers are moving away from fancy, one-time demonstrations and toward machines that could eventually be both helpful and trustworthy. The shift here is from building a tool for a specific task to building an agent that understands the physical logic of the world.
Liked this one? The next lands at breakfast.
Every story in tomorrow's AI news, rebuilt in plain English — five minutes, sources linked, free forever.
By joining you agree to receive Article's daily newsletter — unsubscribe in one click. Privacy