According to a post on Xiaomi's official technology Weibo account, the company has open-sourced its embodied foundation model, Xiaomi-Robotics-1.
Xiaomi Technology stated that this open-source release encompasses the complete workflow from real-world robot post-training to model deployment, and also provides evaluation code for the relevant benchmarks.
It is reported that Xiaomi-Robotics-1 was pre-trained on over 100,000 hours of UMI data and underwent post-training using more than 10,000 hours of cross-embodiment data. On July 16 of this year, Lei Jun posted on Weibo that Xiaomi-Robotics-1 was officially released, noting that the model had preliminarily validated the Scaling Law effect within the field of embodied foundation models. The pre-training on 100,000 hours of real manipulation data and post-training on cross-embodiment data allow different robot bodies to be used out-of-the-box and execute tasks with a single command.
Lei Jun stated that, according to two authoritative embodied intelligence benchmarks, Robocasa365 and RoboDojo, Xiaomi-Robotics-1 ranks first globally overall. He added that Xiaomi's robotics team has recently achieved a series of significant technical milestones, preliminarily completing a comprehensive layout of "hardware, data, and models," and will next drive the advancement of robotics into more open and complex real-world environments.
Comments