X-Square Robot: a foundation stack for general-purpose robots, finally readable

The X-Square Robot laboratory publishes a three-layer "foundation stack" for embedded AI. It's not yet a robot GPT - but the architecture reading, it, is ripe.

In plain terms

The X-Square Robot lab has just published a three-layer architecture to make robots truly versatile: a perception model, an action model, and a control model. Nothing revolutionary taken piece by piece. But the way it's stacked—and argued—marks a step.

The context

Since Google RT-2 (2023), Figure 02 (2024), and DeepMind's Vision-Language-Action (2025), the field is looking for the "foundation model of the robot." Two schools are competing: end-to-end (one big model from pixel to motor) against modular stack (separate perception, planning, and control). End-to-end has mathematical elegance; the stack has engineering and debugging.

The data

According to IEEE Spectrum, X-Square Robot explicitly defends the modular approach, structured in three distinct layers:

• Perception - pre-trained vision-language model.
• Task Planner - decomposition of goals into sequences of actions.
• Control policy - reinforcement learning combined with imitation learning, simulator-to-real transfer.

The article positions the stack as a target for "general-purpose robots." No public third-party benchmark yet.

The analysis

The interest is not in the displayed performance—unverifiable—but in the methodological choice. After three years of "just bigger model," several teams are returning to modularity. Why? Because on the robot, end-to-end latency matters, safety requires explicit interception points, and debugging a failure mode requires isolating the faulty layer. An opaque end-to-end model, when the gripper breaks a glass, leaves you speechless.

So what

For a robotics investor: no longer bet on the promise "a giant model will solve everything." Look for who has a documented, testable, debuggable stack. For an engineer: the real question for 2026-2027 is no longer which model, but which interface between layers. That's where the value will crystallize, just as the protocol did for the web.