China just published the recipe for robot intelligence that Silicon Valley is still workshopping in private Slack channels.
The Summary
- X Square Robot released an open-source "foundation stack" for general-purpose robots, treating physical AI like LLMs: pretrain broad, get general capability
- Core thesis: robots need integrated world models (predict physical changes) and action models (perception + planning + execution), not Frankenstein'd perception-planning-control parts
- Three design principles distinguish this from typical robotics: (1) data unit is interactions that change the world, not just joint trajectories, (2) pretraining yields usable capability not just fine-tuning starting points, (3) behavior modeled around physical events not fixed time slices
The Signal
Robotics has been doing software archaeology for decades. Stitch together a perception module from 2015, a planning layer from 2018, a control system from 2021, and hope they talk to each other long enough to pick up a coffee cup. X Square Robot's stack makes an argument the field has been too fragmented to make: what worked for LLMs, integrated pretraining on broad data yielding transferable capability, should work for robots too.
The technical choices reveal how seriously they took the LLM analogy. Most robotics training data captures "the robot moved its arm like this." X Square captures "the robot moved its arm and the cup relocated." The difference matters because the second teaches cause and effect in physical space, not just motion imitation. Their world model predicts what happens when you act, their action model decides what to do, and both train on the same robot-free interaction data. You can simulate physics without owning hardware.
"The basic unit of robot data is an interaction, not a trajectory."
The open-source commitment is the real wildcard here. Embodied AI has been a walled garden, every lab hoarding data and models like they're sitting on the next GPT moment. X Square is Chinese, which adds geopolitical texture, but they are publishing the recipe. If it works, the timeline for capable home robots compresses. If it doesn't, everyone learns what not to build.
Compare this to what most robotics companies ship: task-specific demos that look impressive in controlled environments and fall apart when the lighting changes. Pretraining for usable capability, not just a head start on fine-tuning, means a robot trained on their stack should handle novel objects and spaces without retraining from scratch. The bet is that physical intelligence generalizes the same way language intelligence does.
Key technical departures:
- Event-based behavior modeling vs. fixed time-step control (matches how humans think about tasks: "when the door opens" not "at timestamp 3.2 seconds")
- World model and action model as independent but complementary families sharing a codebase (parallel training, joint deployment)
- Robot-free pretraining data (simulation and video) feeds both models equally
What makes this a foundation stack and not just another robotics framework is the interdependence. The world model's predictions inform the action model's decisions. The action model's attempts generate data the world model learns from. Neither works at scale without the other, which is why they built and released them together.
The Implication
If X Square's stack becomes the Pytorch of embodied AI, we will look back at 2025 as the year robotics stopped being a hardware problem and became a data problem. The companies that win will be the ones that figure out how to generate interaction data at LLM-training scale. Watch for: (1) open-source forks and extensions from university labs, (2) hardware startups adopting the stack to skip years of software development, (3) benchmarks that test generalization across robot morphologies, not just task success rates.
For technical founders: the actionable move is evaluating whether your robotics work assumes task-specific training or whether it can plug into a pretrained foundation. If you are still hand-tuning controllers, you are building in the wrong layer of the stack.