A walking supercomputing center! 80 robots debuted at the World Humanoid Robot Games: the first fully autonomous group collaboration in human history.

On August 23, the second World Humanoid Robot Games opened at the National Speed Skating Oval in Beijing, where 80 Booster T2 humanoid robots from Accelerated Evolution made a collective appearance at the opening ceremony, delivering a large-scale fully autonomous group coordination performance.

Accelerated Evolution emphasized that there were no marker lines on the field, nor were there operators holding remote controls following behind.

At the scene, the 80 robots sequentially completed formation changes, "writing" the games emblem, event icons for table tennis and badminton, as well as the word "BEIJING" on the ground, before neatly running off the field.

Notably, these actions were not controlled by a centralized system.

Accelerated Evolution stated that all 80 robots are equipped with 2070 TFLOPS of edge-side computing power. Without central command, each robot makes independent decisions, coordinates in real time, and autonomously reaches designated positions to ultimately complete the entire patterned formation.

Calculated at 2070 TFLOPS per unit, the total computing power of 80 T2 robots operating simultaneously is approximately 170,000 TFLOPS, equivalent to a "walking supercomputing center."

According to official introduction, this also marks the first time in human history that nearly a hundred fully autonomous humanoid robots have completed group coordination in the physical world.

In terms of hardware, the Accelerated Evolution T2 stands 1.4 meters tall, features up to 75 degrees of freedom across the body, uses self-developed high-torque joints with a maximum torque of 140Nm, and supports a maximum static load of 10kg on both arms.

The professional edition comes standard with an NVIDIA Thor chip, delivering 2070 TFLOPS of edge-side computing power, paired with a 14-core CPU with a maximum clock speed of 2.6GHz, and 128GB of 256-bit LPDDR5X memory, achieving a minimum edge-side multimodal inference latency of 100ms.