IT Home reports a Xiaomi robot trial in a company apartment.

The big change

  • The reported run spans multiple rooms and household tasks.
  • Earlier research trained the model on manipulation data and aligned actions with different robot bodies and instructions.
  • The trial report does not establish repeat success or intervention rates.

From a reported run to household work

On November 3, IT Home reported Xiaomi’s video describing a system based on Xiaomi-Robotics-1 in trial operation at a company youth apartment. The publication relayed a run of more than 30 minutes across three rooms and over 50 tasks, including tidying surfaces and collecting rubbish.

The report describes a loop joining data collection, model inference and task execution. It does not provide the number of attempts, interventions or a repeated-run success rate. BIG CHANGE could not independently access the underlying video.

How the model was trained and tested

Xiaomi’s separate July paper describes a vision-language-action model trained on real-world manipulation recordings. Post-training aligns its learned actions with different robot bodies and language instructions.

In one earlier evaluation, researchers tested four familiar task types with unseen objects and environments, without task-specific or environment-specific fine-tuning. For its five-billion-parameter comparison, Xiaomi reported 75% overall success after action pre-training on 20,000 hours of recordings, versus 26% without action pre-training. Both used identical post-training. Four familiar task types were tested with unseen objects and environments. That supports a pre-training benefit under those test conditions.

A separate evaluation used four tasks held out from the in-house robot data, then fine-tuned on those tasks and tested each model ten times per task.

A separate household research system

Stanford and Caltech’s HomeBody uses persistent spatial memory and a library of reusable skills to organize longer household tasks. A vision-language model calls skills such as navigation and grasping and receives feedback on their execution. BIG CHANGE previously covered that separate project.