XiaomiRobotics/Xiaomi-Robotics-1

What it solves

Xiaomi-Robotics-1 (XR-1) addresses the "data barrier" in robotics—the scarcity of high-quality, manually labeled robot data that limits the scaling of policy models. It enables robots to perform mobile manipulation in unseen environments and adapt to new, complex tasks using very little demonstration data.

How it works

XR-1 is a Vision-Language-Action (VLA) model that combines a pre-trained Vision-Language Model (Qwen3-VL) with a Diffusion-Transformer (DiT) using a Mixture-of-Transformers (MoT) architecture. It uses a two-stage training process:

1. Pre-training: The model is trained on over 100,000 hours of "embodiment-free" trajectories. An auto-labeling pipeline segments these trajectories and generates descriptions of state transitions, teaching the model general action generation without needing specific robot hardware data.
2. Post-training: The model undergoes alignment using cross-embodiment data (real-robot and open-source data) to map general abilities to specific robot hardware and to align the model with natural-language instructions.

Who it’s for

This project is designed for robotics researchers and developers building foundation models for mobile manipulation and those looking to deploy general-purpose robot policies that can generalize across different environments and tasks.

Highlights

• Massive Scale: Pre-trained on 100K+ hours of real-world manipulation trajectories across 1,700+ scenarios.
• SOTA Performance: Ranks first on RoboCasa365 and RoboDojo leaderboards.
• Data Efficiency: Can learn new complex tasks with only a few hours of demonstrations per task.
• Real-Time Optimized: Uses asynchronous execution to reduce inference latency and is optimized for consumer GPUs.