Xiaomi published the full weights, training pipeline, and deployment code for Xiaomi-Robotics-1, an embodied-AI foundation model built to give robots general physical skills rather than one narrow task. Inside AI News reported the release on August 5, noting that Xiaomi also shipped benchmark evaluation scripts alongside the model, so outside developers can test its claims rather than take the company’s word for them.
That detail matters more than it looks. Xiaomi first showed Xiaomi-Robotics-1 in July as a ready-to-run foundation model. This week’s release is the part that determines whether anyone outside Xiaomi actually builds on it: the code, not the demo.
The training recipe leans on scale rather than a novel architecture. Xiaomi built the model’s base skills on more than 100,000 recorded hours captured through UMI, short for Universal Manipulation Interface. That is a capture method that records human hands performing everyday tasks and retargets that motion onto a robot gripper. A second training pass added upward of 10,000 hours of data gathered across several distinct robot bodies, an attempt to make one model transfer across hardware instead of locking onto a single arm design.
Open weights change less in robotics than they do in language models. A shared checkpoint for a text model can be tested on any laptop in minutes. A robotics checkpoint needs an actual arm, calibrated cameras, and a lab willing to risk hardware on an unproven policy. Xiaomi’s code removes the software barrier to entry. It does nothing about the capital and physical-infrastructure barrier that has kept embodied AI a slower field than chatbots or coding assistants.
The release also leaves real gaps. Xiaomi has not disclosed the model’s parameter count, its architecture, or benchmark results against rival systems, so developers have to run their own tests to learn whether it clears any meaningful bar. The company has also not detailed how it filtered a dataset this large for quality, a gap worth noting because UMI captures are known to be noisy and, in most published academic use, largely confined to tabletop manipulation.
Figure AI and Tesla keep their humanoid software proprietary, arguing that tightly coupled hardware and models let them iterate faster without outside interference. Xiaomi is making the opposite bet: a phone and appliance maker with no flagship humanoid product to defend has more to gain from seeding a developer ecosystem than from guarding a model few outsiders could even run on Xiaomi’s own hardware. Hugging Face’s LeRobot project has spent roughly two years assembling that kind of open robotics community without a training-scale foundation model to anchor it. Xiaomi’s release hands LeRobot’s ecosystem exactly that anchor, whether or not that was the intent.
Robotics teams should treat the open weights as a starting point, not a verified leaderboard entry. Run Xiaomi-Robotics-1 against your own hardware and task set before letting the training-hour figures stand in for real evaluation. The more durable story is a consumer electronics company, not a dedicated robotics lab, choosing to give this pipeline away first.
Reported by Inside AI News on August 5, 2026.
