product updateXiaomi

Xiaomi Humanoid Completes First 8-Hour Car Factory Shift; Xiaomi-Robotics-1 Enters Apartment Trial

TL;DR

Xiaomi says its humanoid robot worked an eight-hour shift at its car factory, covering the full daily production demand of the center console side-cover sorting station. The company also started a home-service trial of its Xiaomi-Robotics-1 embodied foundation model at a Xiaomi Youth Apartment.

3 min read
0

Xiaomi says its humanoid robot has for the first time completed an eight-hour work assignment at its car factory, covering the full daily production demand of the center console side-cover sorting station. The company also began a home-service trial of its Xiaomi-Robotics-1 embodied foundation model at a Xiaomi Youth Apartment, its rental housing for employees.

Xiaomi disclosed both milestones at the 2026 World Internet Conference Wuzhen Summit, where its robotics team received a Leading Technology Award for foundation model technology aimed at large-scale robot manipulation.

Factory: from three hours to a full shift

According to Xiaomi, the robot corrected anomalies on its own during the eight-hour shift and coordinated with production line systems. The company describes this as a preliminary validation of line-level capability.

The shift follows a factory internship that began in March:

  • March: the robot worked three hours autonomously at a self-tapping nut loading station, reaching a 90.2% two-sided installation success rate while meeting the line's 76-second cycle time.
  • July: that success rate reached 98%, and the robot took on two more stations, side-cover sorting and bin folding.
  • Now: Xiaomi puts success rates across the three stations at 94% to 98%.

Home: 50+ tasks in a single run

In the apartment, engineers built an agent system on top of the pretrained Xiaomi-Robotics-1 model, covering data collection, model inference and task execution. They refined it over several rounds of post-training.

In a validation run, the system worked continuously for more than 30 minutes across three rooms and completed more than 50 tasks. These included preparing tools, tidying a sofa and tables, clearing trash, sorting items back into place, inspecting rooms, rechecking the whole unit and leaving. A video released by Xiaomi presents the run as a single 32-minute take.

Xiaomi has not said how many robots are involved in the apartment trial, which apartment hosts it, or how long the trial will run.

The model stack

Xiaomi-Robotics-1 was unveiled on July 16 and open-sourced in August. Xiaomi says the model was:

  • Pretrained on more than 100,000 hours of real-world manipulation trajectories spanning over 1,700 scenes and 260 task types.
  • Post-trained on more than 10,000 hours of cross-embodiment data so it can be adapted to mobile manipulators and dual-arm robots.

The company says it ranked first overall on the RoboCasa 365 and RoboDojo leaderboards in July. Its separate vision-language-action model, Xiaomi-Robotics-0, uses a hybrid architecture that Xiaomi says reaches 80-millisecond inference latency on a consumer graphics card.

Parameter count, context window and pricing for these models are not part of the announcement.

What this means

The factory result is a duration milestone: the robot went from a three-hour autonomous run at one station in March to covering a full day of demand at the side-cover sorting station. Xiaomi frames it as preliminary validation, not a finished line deployment, and the success rates and shift results are company-reported.

The apartment run is a validation, not a deployment. Its scope is a single run of more than 30 minutes across three rooms, and Xiaomi has given no details on trial size or length.

The notable claim is architectural: Xiaomi presents the factory and the home as one model stack, with Xiaomi-Robotics-1 as the pretrained base and task-specific post-training and agent systems on top. Because the model is open-sourced, the pretraining and cross-embodiment data figures Xiaomi cites describe a base others can adapt to mobile manipulators and dual-arm robots.

Related Articles

model release

Xiaomi open-sources Xiaomi-OCR-0, a 0.8B document model scoring 96.83 on OmniDocBench v1.6

Xiaomi has released the weights and inference code for Xiaomi-OCR-0, a 0.8B-parameter model that handles document parsing, key information extraction and OCR visual question answering. Xiaomi reports 96.83 on OmniDocBench v1.6, first among models under 1B parameters and second overall.

product update

Fireworks claims up to 1,500 tokens/sec on DeepSeek V4.1 Flash in Ultra-fast inference preview

Fireworks announced a set of enterprise AI products at its first annual Forge conference on November 3, 2026. The lead item is an Ultra-fast inference preview that the company says reached up to 1,500 tokens per second on DeepSeek V4.1 Flash. It also announced a private preview of Fireworks Sovereign for customer-owned infrastructure.

product update

Xcode 27.3 beta 1 adds GitHub Copilot to Apple's built-in coding assistant

Apple released Xcode 27.3 beta 1 on November 3, 2026, adding GitHub Copilot to the Xcode coding assistant. The beta also includes Swift 6.4, Device Hub improvements and UI testing support in Swift Testing. Xcode 27.3 is expected in December, according to 9to5Mac.

product update

Moonshot AI's Kimi launches game development package with K3.1 agents inside Unity, Unreal Engine and Blender

Moonshot AI has released a game development and creative work solution built on the Kimi K3.1 model series, Kimi Code, Kimi Work and plugins for Unity, Unreal Engine and Blender. Kimi says one Unity demonstration ran for 20 hours without human intervention.

Comments

Loading...