Embodied Intelligence: From Data Scarcity to Real-World Robotic Manipulation – JD Explore Academy’s System Architecture and Research Advances

Large‑model breakthroughs have sparked a wave of embodied‑intelligence research, with robots moving from laboratory showcases—like Boston Dynamics’ dancing machines and Tesla’s Optimus—to everyday applications that can perform tasks such as dancing, kung‑fu, and eventually assist humans in daily life.

The field focuses on three functional blocks: motion, operation, and navigation. While motion (e.g., dancing) demonstrates macro‑level control, the real challenge lies in precise operations like picking objects, washing clothes, or cooking, which require robust perception, intent understanding, and millimetric tactile feedback.

JD Explore Academy built a highly extensible embodied‑intelligence system on the ROS platform. A central scheduler coordinates navigation, perception, task planning, and remote control modules, while asynchronous model inference, gRPC communication, and a parent‑child routing mechanism address latency and speed issues. This architecture successfully powered a coffee‑making robot with an 80% success rate in real‑world scenarios.

For dual‑arm dexterous hands, the team developed a high‑frequency (≥50 Hz) integrated tele‑operation solution that fuses inertial and visual motion capture, enabling lightweight, low‑cost, and highly extensible control. The system achieves sub‑50 ms latency, allowing precise replication of human arm‑hand motions.

To overcome data scarcity, a novel end‑effector imitation method was proposed. It learns unified operation trajectories through object‑centric visual feature extraction, pose estimation, and strategy learning, achieving over 90% success in coffee‑making and more than 50% improvement on tasks such as barcode‑gun grasping, doll‑picking, and box‑moving.

In 2023, JD Explore Academy released China’s first dual‑arm mobile‑robot operation dataset, JD ManiData, and an open‑source atomic‑skill library built on a three‑wheel data‑driven framework, addressing the current data‑deficiency problem in embodied intelligence.

Future work will continue to integrate vision‑language‑action large models, pre‑training, and reinforcement learning to boost operation success rates and generalization, expanding applications to service, guide, household, and parcel‑delivery robots.