Gaode Robot Dog Crosses Streets, Buys Milk Tea

Recently, several social media users posted videos showing a robot dog, branded with the Gaode Maps (Amap) logo, autonomously crossing streets and buying milk tea outside Gaode’s headquarters. The footage confirms that this is indeed Gaode’s own robot dog.

According to a Gaode representative, the company has been making deep investments in embodied intelligence and is actively exploring hardware forms such as quadruped robots and humanoid robots. Its first quadruped robot is expected to be officially released soon.

Videos captured by bystanders reveal that the robot dog is capable of independently navigating urban environments. When crossing the street, it waits for the traffic light to turn green before proceeding. While moving forward, it can detect pedestrians and actively avoid them.

In another clip, the robot dog enters a milk tea shop on its own, communicates the pickup number to the staff, receives the drink, and then exits the store autonomously before returning to Gaode’s headquarters.

Industry observers note that for a robot dog to operate independently in open, real-world environments, it must possess highly advanced capabilities in perception, decision-making, and motion control. Gaode’s long-standing accumulation of massive spatiotemporal data may provide an ideal foundation for training such capabilities. This data advantage is likely a key reason behind the company’s confidence in entering the embodied intelligence sector.

Public information shows that Gaode established its Embodied Intelligence Division in January this year. Since then, it has released multiple embodied AI models, achieving state-of-the-art (SOTA) performance across more than ten authoritative benchmarks.

In February, Gaode introduced two foundation models in its ABot series: the embodied manipulation foundation model ABot-M0 and the embodied navigation foundation model ABot-N0.

Notably, ABot-N0 is the first model to unify five core navigation tasks within a single system, breaking away from the traditional “one task, one model” paradigm. The model has already set new records across seven international benchmarks, including CityWalker, SocNav, R2R-CE, RxR-CE, HM3D-OVON, BridgeNav, and EVT-Bench.

In addition, Gaode’s Abot-world series world models have recently topped two major international evaluations: the Agibot World Challenge and World Arena.

The Agibot Challenge, initiated by Agibot Robotics and supported by leading academic conferences, focuses on system-level challenges that close the loop from language instructions to real-world actions, emphasizing executability and physical plausibility. Meanwhile, World Arena aims to establish standardized evaluation systems for world models, driving the transition from “perceptual intelligence” to “cognitive intelligence.”

The Abot-world models demonstrate strong capabilities in internalizing physical laws and performing long-horizon dynamic predictions. Unlike many models that generate only short, static, or decorative video clips, Abot-world can accurately anticipate object trajectories in complex interactions—such as sliding, tipping, stacking, and fluid changes—while maintaining consistent multi-step causal logic.

Experts believe that this “reasoning-capable” generative ability gives the model practical value in task planning, anomaly prediction, and autonomous decision-making.

With its combination of real-world robotics demonstrations and advanced AI models, Gaode appears to be positioning itself as a serious contender in the embodied intelligence race. The robot dog’s ability to navigate city streets and complete everyday tasks like picking up milk tea offers a glimpse into a future where intelligent machines seamlessly integrate into daily life.