Recently, the JD Cloud JoyBuilder model development platform has witnessed a comprehensive upgrade. It now seamlessly supports thousand-card training tasks for the GR00T N1.5 model, marking its status as the inaugural AI development platform within the industry to embrace the LeRobot open-source training framework for thousand-card-scale embodied AI. When pitted against the open-source community version, this platform's training efficiency has soared by 3.5 times. Specifically, the training duration for thousand-card tasks encompassing over 100 million data points has been drastically slashed from 15 hours to a mere 22 minutes.

With a laser focus on embodied AI model training, the JD Cloud AI Infra and its associated teams have embarked on full-stack optimizations centered around this platform. These optimizations span pivotal domains such as embodied data pipelines, model computation, and underlying infrastructure. Moreover, the platform has been fine-tuned to align with the latest LeRobot training data protocol, thereby offering accelerated backing for the widespread deployment of embodied AI.