WAIC 2026: TARS with AWE 3.5 Breaks the "Deep Water" of Industry — Workaholic Robots and Built-in Physical Simulator
Chinese company TARS (它石智航) showcased at WAIC 2026 a new version of its embedded model AWE 3.5 — the first embedded model supporting the full cycle from pre-training to post-training. The demonstration includes numerous tasks (assembly, packaging, sorting) performed by a single model without switching parameters, as well as an interactive physical simulator for reinforcement learning in a virtual environment.
At the WAIC 2026 conference, TARS (它石智航) demonstrated a fully automated production line for assembling automotive wire harnesses, where multiple robots work collaboratively. The core of the line is the new AWE 3.5 model—the first embedded model that combines pre-training and post-training within a single framework. The model can perform multiple tasks without reloading or switching parameters, ranging from industrial assembly to phone packaging and part sorting. A key feature of AWE 3.5 is its built-in physical simulator, trained on real-world data, which allows the model to predict the consequences of actions (for example, the force required to tighten a zipper) and conduct reinforcement learning within its internal world. The model was trained on over one million hours of human-centric data. TARS claims that the time to master a new task has been reduced to just a few hours. The company also emphasizes the importance of long-term memory and spatial understanding: video generated by the model maintains environmental stability over several minutes of interaction. TARS expects that by mid-2027, clear evaluation criteria for embedded AI will emerge, and the industry will shift from demo dances to real work.
Source: QbitAI 量子位 —
original