TARS presented its “Trustworthy Physical AI” strategy at the World Artificial Intelligence Conference (WAIC) 2026 in Shanghai, where its AWE (AI World Engine) embodied foundation model received the SAIL (Superior AI Leader) Award in recognition of its technical innovation and industrial potential. During the conference’s main forum, TARS founder and chief executive officer Dr. Chen Yilun introduced AWE 3.5, the company’s latest embodied AI foundation model.
According to the company, the model was developed using more than one million hours of human-centric real-world data collected and validated in industrial environments. TARS said AWE 3.5 combines action, perception, geometry and tactile sensing within a unified framework and is designed to improve task execution, complex-task performance and long-duration closed-loop interaction. The company also described it as the first embodied-native model to implement a combined pre-training and post-training development approach intended to support continuous iteration.
The company demonstrated robots powered by AWE performing tasks including packing mobile phones, organizing backpacks and sorting screws. TARS said these demonstrations were intended to illustrate the model’s ability to apply physical understanding to practical tasks and adapt to unfamiliar objects and environments. The company stated that it plans to expand its training dataset from more than one million hours to 10 million hours by the end of 2026 to improve generalization, long-horizon reasoning and task execution.
TARS also presented a full-scale automotive wiring-harness production line using multiple A1 robots to grasp, route, connect and assemble flexible wiring harnesses. The production line was featured in WAIC’s “Model Era • Partner City” embodied robotics exhibition within the “Smart Manufacturing Hub.” The company said it is working with Shanghai’s Jiading District and industry partners to validate the technology in industrial settings and advance the deployment of a thousand-unit-scale industrial embodied robot cluster.
The conference also marked the debut of DexHand, a robotic hand mounted on an A1 robot. In a live demonstration with magician Deng Nanzi, the system performed card spreading and shuffling, handwriting and Rubik’s Cube solving. TARS said the demonstration highlighted the system’s real-time perception, precision control and human-robot collaboration capabilities.
