JD Cloud JoyBuilder model development platform upgraded, enhancing embodied intelligence model training efficiency by 3.5 times

JD.com (09618.HK) subsidiary JD Cloud's JoyBuilder model development platform has recently undergone a major upgrade, successfully supporting the industry's top model GR00T N1.5 for kilocalorie training, becoming the first AI development platform in the industry to support the embodied intelligence kilocalorie-level LeRobot open-source training framework, with training efficiency improved by 3.5 times compared to the open-source community version. Based on deep optimization of hardware and software and breakthroughs at the algorithm level, the model training efficiency and stability have been significantly enhanced, reducing the kilocalorie training time for over 100 million data points from 15 hours to 22 minutes, accelerating the scaling of embodied intelligence.

According to reports, regarding the training of embodied intelligence models, JD Cloud AI Infra and related teams have conducted full-stack optimization based on the JoyBuilder model development platform.

In terms of computational optimization for embodied models, extreme optimizations have been made from multiple angles, including the Attention layer, token pruning, and post-training quantization, targeting the computational characteristics of mainstream open-source VLA (Vision-Language-Action) models, comprehensively enhancing the training efficiency of the models.

Additionally, the JD Cloud JoyBuilder model development platform, based on comprehensive optimizations in end-to-end data processing, model computational efficiency, and AI infrastructure, supports the latest protocols for training data of the industry's most mainstream LeRobot and has become the first AI development platform in the industry to support kilocalorie-level open-source training frameworks for embodied models