Tashi AI Releases AWE3.7 Embodied Model with Cross-Domain Generalization

Itshizhihang recently released the general-purpose embodied large model AWE3.7 (AI World Engine) and within ten days published a series of task demonstrations. The core claim is that the same 'embodied brain' can handle industrial production lines, logistics operations, life services, and mobile manipulation, addressing the industry's challenge of 'generalization' in embodied intelligence.

According to the release, AWE3.7 is based on a closed-loop methodology of 'native architecture, dual-prior pretraining, world-model-driven post-training, large-scale validation, and data feedback', leveraging over a million hours of human-centric real data and rich visual-tactile information. These statements are self-reported by the manufacturer and have not yet been independently verified by third parties.

In industrial scenarios, AWE3.7 is claimed to achieve production-line-level performance. Parts sorting tasks require the robot to distinguish tiny differences among randomly scattered screws and nuts, and to autonomously correct errors when disturbed; tape wrapping involves flexible materials, requiring bimanual coordination and real-time trajectory correction to ensure the tape adheres without breaking.

Network cable insertion requires millimeter-level precision and resistance to disturbance; phone boxing is a long-horizon operation that must complete the entire packaging process under varying lighting conditions; screw tightening relies on torque feedback for real-time adjustment, achieving 'accurate alignment, stable tightening, and controlled stopping'. These descriptions come from official demonstrations and do not provide quantitative success rates or error data.

Dynamic scenarios test the robot's real-time perception and operational continuity. Facing soft toys moving on a conveyor belt, the robot must follow the pace to complete bagging; block sorting requires understanding the 'color-target box' rule and dynamically adjusting trajectories when the storage box is moved.

In larger-space tasks, the robot must achieve high integration of mobility and manipulation, executing end-to-end seamlessly without human intervention. These demonstrations showcase AWE3.7's potential in dynamic environments, but they also lack external evaluation.

Home scenarios are seen as a litmus test for general capability transfer. Folding clothes requires bending down to pick up items and placing them in a storage basket, demonstrating loco-manipulation whole-body control; wiping a whiteboard requires both left and right hands to perform the task and can switch hands mid-task, with appropriate force.

Packing a school bag involves multi-object recognition and sequential planning; preparing breakfast requires retrieving utensils from a cabinet and placing them; making a smoothie requires grasping a soft plastic cup, aligning it with the ice outlet, observing the capacity, and stopping in time. These long-horizon tasks demonstrate the model's adaptability in unstructured environments, but the editing and selection of demonstration videos may affect their persuasiveness.

Itshizhihang mentioned that in March 2026, the Itshizhihang A1 robot powered by the AWE model set a Guinness World Record for 'most times assembling sub-millimeter wire harnesses in one hour'; at the same year's WAIC, the AWE general-purpose embodied large model won the SAIL Star, becoming the only embodied foundation model to receive this award.

These honors belong to the AWE series, not specifically to AWE3.7. The release of AWE3.7 itself lacks third-party benchmark testing or peer review, and its 'generalization' capability currently relies mainly on manufacturer claims and demonstration videos.