• 2 min read
Xiaomi says Robotics-1 learned from 100,000 hours
Xiaomi unveiled Robotics-1, its first foundation model for embodied AI and robotics, trained on 100,000 hours of real-world data and set to be open source.
Image: ITzine
Xiaomi has introduced Xiaomi-Robotics-1, its first foundation model for embodied AI and robotics. The company says the system was trained on 100,000 hours of real-world data and is designed for more than a single lab setup, targeting different types of robots instead.
According to Xiaomi, the model is built to handle mobile control and adapt to new conditions without lengthy tuning for each hardware platform. The company first pretrained the system on large datasets, then fine-tuned it across several robotic platforms.
Xiaomi says Robotics-1 ranked first in the RoboCasa365 and RoboDojo tests. Another notable move is the company’s plan to release both the model’s source code and weights, giving developers access to more than a demo.
Xiaomi is not the first company in this area. In 2023, Google DeepMind introduced RT-2, a model designed to transfer knowledge from the internet into robot actions. In 2024, NVIDIA announced Project GR00T for humanoid machines, and researchers from Stanford, Berkeley, and Google DeepMind released the open OpenVLA model.
Recommended reading
Yandex certifies all Alice AI models to ISO/IEC 42001
Xiaomi’s pitch is different: a focus on real-world data and openness. If the company publishes the weights soon, as indicated, the robotics market will get another notable open-source tool.
AI Editor
Ava covers the rapidly evolving world of artificial intelligence, from foundational models and research labs to the real-world economics of intelligence. With a background in computational linguistics, she cuts through the hype to find out what actually works. She firmly believes that benchmarks are just marketing until reproduced in the wild.
via ITzine
