88Z AI, 27 April — Xiaomi's Robotics Division has open-sourced the full real-device post-training pipeline for Xiaomi-Robotics-0. Built on a pre-trained base, the model achieves high-precision tasks like earphone storage with just 20 hours of task data. The team employed strategies such as asynchronous inference, action prefixes with adaptive weighting, Lambda-type masking, and random masking to address the model's 'laziness effect,' balancing action coherence and visual sensitivity.
At a glance
- Open-sourced real-device post-training pipeline enables high-precision operations with 20 hours of data
- Multiple strategies address model laziness, balancing action coherence and visual sensitivity
The update
Xiaomi's Robotics Division has officially open-sourced the full real-device post-training pipeline for the Xiaomi-Robotics-0 large model. Built upon a pre-trained foundation, the system leverages data efficiently, enabling high-precision complex operations such as earphone storage with only 20 hours of task data. To address the potential 'laziness effect' in long-sequence tasks, the research team introduced multiple optimization strategies. These include asynchronous inference mechanisms, action prefixes combined with adaptive weighting techniques, and a triple strategy combination of Lambda-type masking and random masking. These technical approaches work together to successfully achieve a strong balance between robot action coherence and visual perception sensitivity, enhancing overall execution performance.
Sources & editorial note
This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
- 小米技术Source published
Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.