Tsinghua-Harvard Team's Acorn Robot Develops 'Zero-Data' Robot That Learns Through Instinct, Not Training Data

A robotics startup co-founded by researchers from Tsinghua University and Harvard University has unveiled what it calls a "zero-data" robot — a machine capable of learning physical manipulation tasks without any pre-collected training data, demonstration trajectories, or visual models. Acorn Robot, founded by Dr. Jiang Yao — who holds a PhD in Mechanical Engineering from Tsinghua and completed a postdoctoral fellowship in Neuroscience at Harvard — has developed a robot that uses an onboard autonomous decision model called "Natus," which stands for "instinct-driven behavioral emergence." The system learns entirely through physical trial and error, attempting tasks, failing, and self-correcting in real time. The robot's hardware is strikingly minimal. It uses only a simple one-degree-of-freedom industrial parallel gripper with two wedge-shaped black jaws equipped with tactile sensors on their inner surfaces. There is no external camera, no cloud-based AI, and no training data pipeline. Despite these limitations, the robot can perform surprisingly dexterous tasks. In a demonstration, the robot picks up a credit card lying flat on a table — a notoriously difficult task for conventional industrial grippers. It accomplishes this by wedging one jaw against the card's edge and using the table surface as a fulcrum to lift it. The system typically requires eight to nine attempts before discovering an effective strategy through self-guided experimentation. According to the company, the technology has already passed the proof-of-concept stage at one of China's top cosmetics companies and achieved scaled deployment. Acorn Robot is targeting B-end flexible manufacturing scenarios where adaptability — not data volume — is the critical bottleneck. Dr. Jiang argues that current mainstream embodied AI approaches, including vision-language-action models, world models, and simulation-based learning, face fundamental problems in real physical interaction. Unpredictable contact forces and hardware differences between individual robots make data-driven approaches an "unfillable bottomless pit," he contends. "There is no universally best model," Jiang said. "Only the model best adapted to a specific robot." The zero-data approach represents a philosophical departure from the data-hungry paradigm that dominates much of contemporary robotics AI, suggesting that instinct-driven learning on physical hardware may be a more practical path toward general-purpose robotic manipulation.