WAIC 2026: TARS with AWE 3.5 Breaks Through Industry "Deep Water" — Workaholic Robots and Built-in Physics Simulator

At WAIC 2026, TARS (Tashi Intelligence) unveiled its embodied native foundation model AWE 3.5, capable of multitasking without reloading. The model powers a fully automated wiring harness production line and includes a built-in physics simulator for reinforcement learning in a digital twin, marking a breakthrough in embodied scaling.

At WAIC 2026, Tashi Intelligence (TARS) demonstrated a fully automated wiring harness assembly line using multiple robots coordinated by their new AWE 3.5 model. The model is a native embodied foundation that unifies vision, language, and action from pretraining, enabling robots to perform diverse tasks like packaging, plugging cables, and parts sorting without switching parameters. AWE 3.5 also includes a world model that simulates physics, allowing reinforcement learning in a simulated environment before real-world execution. The model is trained on over one million hours of human-centric data, and the company claims that new tasks now require only hours of data collection. TARS chief scientist Ding Wenchao stated that embodied scaling has fully released, and a turning point may appear in the first half of 2027. The company is also working on dexterous hands for high-precision tasks and expects significant industrial trends within 12-18 months.

Сокращения: AWE = Agile World Engine — Agile World Engine (название модели); WAIC = World Artificial Intelligence Conference — Всемирная конференция по искусственному интеллекту; VLA = Vision-Language-Action — Зрение-Язык-Действие; SFT = Supervised Fine-Tuning — Контролируемая донастройка; Infra = Infrastructure — Инфраструктура

Source: QbitAI 量子位 — original