Tashi Intelligence Debuts AWE 3.5 Embodied Native Brain at WAIC, Demonstrating Multi-Task Industrial Prowess

Tashi Intelligence (它石智航) made its WAIC debut at the 2026 World Artificial Intelligence Conference in Shanghai, showcasing the AWE 3.5 embodied native brain in live demonstrations. The robot acted as an industrious worker, seamlessly switching between desktop organizing, phone packing, network cable plugging, and parts sorting—all powered by a single base model without reloading parameters.

The core innovation of AWE 3.5 is its unified architecture. Unlike conventional approaches where different tasks require different model configurations, AWE 3.5 integrates all capabilities into one model. On the exhibition floor, the robot handled whatever task was presented, demonstrating remarkable generalization.

Technically, AWE 3.5 trains vision, language, and action modalities jointly from the pre-training stage, using the same architecture for both action output and future prediction. This differs from the popular VLA (Vision-Language-Action) approach, where vision and language are coupled in pre-training but action is only introduced later through supervised fine-tuning.

The architecture consists of two main components: the Base Policy, which formulates strategies and issues commands, and the Action Condition World Model, which predicts environmental changes based on actions. Together they form a complete reinforcement learning loop, simulating executions in the model's 'mind' before committing to physical actions—an order of magnitude faster than trial-and-error on real hardware.

On the data front, Tashi Intelligence has accumulated millions of hours of human-centric data. AWE 3.5's model structure is specifically strengthened for long-term memory and temporal-spatial understanding, maintaining environmental stability during multi-minute continuous interaction loops. Post-training can extract multiple action segments from continuous processes without worrying about environmental collapse.

The company is also developing its own dexterous hands using a top-down approach: first identifying human actions and matching human data, then defining hardware design in reverse. While grippers can solve about 80% of problems, the remaining 20% requires dexterous hands. Whether these hands can successfully pass commercial validation may become a key milestone for the embodied AI industry.

Tashi Intelligence acknowledges that the embodied AI landscape remains fragmented with no company yet producing a GPT-3.5-level breakthrough. A watershed moment may arrive between late 2026 and 2027, when the true test shifts from single-scenario tasks to reliable multi-scenario deployment—a trajectory similar to autonomous driving's evolution from demo routes to large-scale operations.