Xiaomi open-sources XR-1 robot AI trained on 1 lakh hours of real-world data

New Delhi: Xiaomi has open-sourced Xiaomi-Robotics-1, also called XR-1, a robot foundation model built to help machines understand instructions and complete physical tasks. The release includes the training pipeline, deployment tools and benchmark evaluation code.

The model was pre-trained on more than 1 lakh hours of real-world manipulation data collected across over 1,700 settings. Xiaomi then used more than 10,000 hours of data from different robot systems for post-training. The company said, “Xiaomi-Robotics-1 is now open source!”

Xiaomi-Robotics-1 is now open source!🎉

Xiaomi-Robotics-1 was pre-trained on more than 100,000 hours of UMI data and post-trained with over 10,000 hours of cross-embodiment data.

This open-source release covers the complete pipeline from real-robot post-training to model… pic.twitter.com/asI6zwWBXe

— Xiaomi Tech (@XiaomiTech_) August 5, 2026

What Xiaomi-Robotics-1 can do

XR-1 is a Vision-Language-Action model. This means it can study what a robot sees, understand a written or spoken instruction, and then decide which physical action should follow.

The model is meant for mobile robots working in places they have not seen before. Xiaomi said it can handle unfamiliar objects and complete tasks without being trained separately for every room or machine.

Its training data covered:

• Homes and kitchens
• Shops and commercial spaces
• Factories and industrial sites
• Outdoor locations

The model combines Qwen3-VL with a Diffusion Transformer using a Mixture-of-Transformers setup. Xiaomi said this design helps it process instructions while keeping response time low enough for real robot use.

Trained on homes, factories and daily tasks

For post-training, Xiaomi collected more than 7,200 hours of real robot data inside homes. Tasks included arranging sofas, sorting shoes and putting away kitchen items.

XR-1 can also learn new jobs from a small amount of demonstration data. Xiaomi tested tasks such as packing phones, refilling printers, loading laundry and packing boxes.

With less than 10 hours of demonstrations per task on average, XR-1 recorded an overall success rate of 75 per cent. The figure rose to 85 per cent with less than 40 hours of training per task.

Open-source release targets robot developers

Xiaomi has released the model through Hugging Face. The software is built around the Hugging Face Transformers system and is optimised for consumer GPUs.

The company has included tools covering real-robot post-training, model deployment and benchmark testing. This gives researchers and developers access to more than just the model files.

Xiaomi said XR-1 recorded the highest scores across four simulation benchmarks, including RoboCasa and RoboDojo. These are company-reported results and will now face wider testing after the open-source release.