RoboScience

RoboScience was founded in December 2024 by Tian Ye (田野), former Apple AI Platform technical lead who spent 7 years at Apple building their machine learning platform (described internally as "Apple's PyTorch and CUDA") powering camera, Siri, and Apple Intelligence. Tian Ye earned his bachelor's degree in Physics (ranked 1st) from USTC and a master's degree in Computational and Mathematical Engineering from Stanford AI Lab under Andrew Ng. The company is headquartered in Shenzhen Qianhai. In May 2026, RoboScience completed a 1 billion RMB (approx. $140M) Series A funding round, the largest of its kind in the embodied AI space that month. RoboScience's core technical contribution is the Object Trajectory representation — a unified 3D object-level state representation that captures how objects change over time (position, orientation, contact, deformation). This serves as an "embodied token" that decouples perception and world modeling from robot-specific control, enabling cross-embodiment generalization. Under this framework, the system defines what should happen to objects (object state changes) rather than how a specific robot should move. The company's architecture is called VLOA (Vision-Language-Object-Action), built around their Visics general-purpose embodied AI model — combining an embodied world model (pre-trained on millions of hours of internet video data) with a general-purpose manipulation model (trained on hundreds of billions of manipulation trajectories generated by their proprietary differentiable physics engine RoboMirage). This data pipeline achieves data costs at 1/20 to 1/200 of traditional real-robot collection. Chief Scientist Shao Lin (邵林) is an Assistant Professor at NUS Computer Science, holding a PhD from Stanford advised by Jeannette Bohg and Turing Award laureate Leonidas J. Guibas. His team won the ICRA 2025 Best Paper Award in Robotic Manipulation and Motion, and received an ICRA 2026 Best Paper Award nomination. Key publications include D(R,O) Grasp, UniGrasp, and SAM-RL. The company's approach fundamentally redefines embodied learning from action imitation to object-state-driven task completion, positioning Object Trajectory as a universal interface that any robot hardware can consume — directly addressing the cross-embodiment generalization challenge that has limited end-to-end imitation learning approaches.