RoboScience Unveils Visics, a General-Purpose Embodied AI Model

Beijing-based embodied intelligence company RoboScience officially unveiled its general-purpose embodied AI model Visics on June 24, complete with a full technical disclosure of the VLOA (Vision-Language-Object-Action) architecture. The company demonstrated real-world applications including furniture assembly, dexterous grasping, and dynamic assembly line operations using its proprietary technology. The VLOA architecture introduces a novel unified intermediate representation standard called Object Trajectory (3D point cloud trajectory), establishing a layered and decoupled framework centered on object manipulation. Unlike mainstream approaches that train models on specific robotic joint trajectories tied to particular hardware configurations — learning "how a gripper picks up a cup" rather than understanding the concept of grasping itself — Visics separates the cognitive and execution layers, enabling true cross-platform generalization across different robot platforms, object types, and task scenarios. Visics operates on a dual-engine architecture: an Embodied World Model trained on massive internet video data to learn object physics, motion patterns, and force dynamics, and a General Operation Model that translates object trajectories into hardware-agnostic control commands. The two engines communicate through the VLOA framework, with Object Trajectory serving as the unified interface between perception and action. This decoupled design achieves generalization across three critical dimensions: adapting to any robot platform, manipulating any object type (rigid, articulated, or soft deformable), and autonomously completing diverse tasks. RoboScience has also developed a proprietary high-precision simulation engine, RoboMirage, combined with automated video annotation pipelines to generate training data at 1/20 to 1/200 the cost of traditional approaches, with expansion to over 1TB of manipulation trajectory data planned by end of 2026. Backed by investors including JD.com, SenseTime, Fortune Capital, CMB Capital, and Sinovation Ventures, RoboScience operates R&D and production centers across Beijing, Shenzhen, Suzhou, and Hangzhou. The company is piloting deployments with retail, logistics, and elderly care enterprises, with plans to commence standardized robot mass production for industrial and commercial applications later this year.