TARS, the physical AI company founded by Dr. Chen Yilun, has been awarded the prestigious SAIL (Superior AI Leader) Award at the 2026 World Artificial Intelligence Conference (WAIC) in Shanghai. The award recognizes the technical innovation and industrial potential of the company's AWE 3.5 embodied foundation model, which was showcased alongside live robot demonstrations throughout the conference.
The SAIL Award is one of the most recognized honors in the Chinese AI industry, awarded annually to projects that demonstrate exceptional technical advancement alongside real-world application potential.
AWE 3.5: Unified Embodied-Native Architecture
At the WAIC main forum, TARS Founder and CEO Dr. Chen Yilun introduced AWE 3.5, the company's latest embodied-native foundation model. Built on more than one million hours of human-centric, real-world data validated in industrial settings, AWE 3.5 integrates action, perception, geometry, and tactile sensing within a unified framework — moving beyond the modular, assemble-from-parts approach that characterizes many current embodied AI systems.
Compared with the previous Pi 0.5 version, AWE 3.5 approximately doubles task-execution efficiency, improves complex-task performance, and maintains closed-loop interaction across tasks lasting several minutes. The model represents a meaningful step forward in the long-standing challenge of getting AI systems to perform sustained, multi-step physical tasks reliably.
Pre-Training + Post-Training: A New Paradigm
Perhaps most notably, AWE 3.5 is the first embodied model to complete an embodied-native "pre-training + post-training" paradigm designed to support reproducible, scalable, and continuous iteration. This approach mirrors the pre-training/fine-tuning paradigm that has driven progress in large language models, but adapted for the unique challenges of physical AI — where data is multimodal, expensive to collect, and must account for real-world physics and contact dynamics.
By establishing a clear separation between general-purpose pre-training and task-specific post-training, TARS aims to create a foundation that can be efficiently adapted to new domains and robot platforms without requiring full retraining from scratch.
Live Demonstrations at WAIC
During the conference, TARS demonstrated AWE-powered robots performing several practical tasks across its exhibition booth:
- Phone packing: Robots autonomously packed smartphones into boxes with protective inserts, demonstrating fine manipulation and object handling in a logistics-relevant workflow.
- Backpack organization: The robots organized items inside backpacks, handling flexible materials and irregular shapes — a benchmark for everyday assistive robotics.
- Precision screw sorting: Robots identified and sorted small screws by type, demonstrating the visual acuity and tactile precision needed for assembly tasks.
Beyond individual task demonstrations, TARS also recreated a full-scale circular automotive wiring-harness production line using multiple A1 robots. The robots grasped, routed, connected, and assembled flexible wiring harnesses — a scenario that was selected for WAIC's official embodied-robotics exhibition zone as a featured display in the "Smart Manufacturing Hub." The demonstration highlighted the technology's stability, reliability, and delivery value in advanced manufacturing settings.
10 Million Hours by End of 2026
Having surpassed one million hours of high-quality human-centric data, TARS has set an ambitious target: expanding its pre-training dataset to 10 million hours by the end of 2026. If achieved, this tenfold increase in training data volume would further strengthen the model's generalization capabilities, long-horizon reasoning, and real-task execution performance.
The data expansion strategy reflects a broader industry consensus that data scale and diversity are primary drivers of capability improvement in physical AI, analogous to how text corpus scaling drove advances in large language models over the past decade.
Industry Significance
The SAIL Award recognition for AWE 3.5 underscores the growing maturity of China's embodied AI ecosystem. As the physical AI field moves from research demonstrations toward commercial deployment, models that can generalize across tasks, operate reliably in unstructured environments, and be efficiently adapted to new use cases will become increasingly valuable. TARS's embodied-native architecture and its pre-training + post-training paradigm represent one approach to achieving that scalability — and the award suggests the technical community is taking notice.
