Recently, the JD Cloud JoyBuilder model development platform has undergone a key upgrade, successfully supporting the training of the industry-leading model GR00T N1.5 on a thousand-card level.

This move makes JoyBuilder the first AI development platform in the industry to support the LeRobot open-source training framework for embodied intelligence, and it has achieved a significant leap in training efficiency, with a 3.5 times improvement over the open-source community version. Through deep optimization of software and hardware and breakthroughs at the algorithm level, the JoyBuilder platform has significantly improved model training efficiency and stability, reducing the time required for a thousand-card training with over 100 million data points from 15 hours to just 22 minutes, greatly accelerating the process of embodied intelligence moving towards large-scale deployment.

Artificial Intelligence, Robots

To achieve this efficiency improvement, the JD Cloud AI Infra and related teams have carried out full-stack optimization on JoyBuilder around embodied intelligence model training. In terms of embodied data chain optimization, the platform has restructured the data preprocessing and loading process, achieving asynchronous execution of CPU data processing and GPU computing, effectively reducing waiting time; for massive small data files in embodied intelligence, the self-developed high-performance parallel file system, Yuhai JPFS, provides a read bandwidth of over 400 GB/s on a 1024-card cluster through distributed metadata management and intelligent pre-fetching, ensuring continuous high-speed data supply.

In terms of embodied model computation optimization, the team has made extreme optimizations from multiple aspects such as Attention layers, Token trimming, and post-training quantization for mainstream VLA (Vision-Language-Action) models. Additionally, on the infrastructure for embodied models, the platform has built a 3.2T RDMA backend network, ensuring high throughput and low latency for collective communication between thousands of cards through multi-track optimization, topology-aware scheduling, and intelligent oscillation suppression, supporting stable long-term training operations. It also optimizes data scheduling and pipeline through cloud-native AI data lakes, improving end-to-end processing efficiency.

Through full-chain optimization, the JoyBuilder platform supports the latest protocol of the most mainstream LeRobot training data in the industry, establishing its leading position in the field of embodied intelligence AI development platforms.