Fudan Team Develops "Danbao" Robot, Pure-Vision On-Device Brain Enables Autonomous Walking
"Danbao," a robot jointly developed by Fudan University's School of Future Information Innovation and the Jimu Tianshu team, was showcased at the 26th China International Industry Fair. Without preloaded maps and without remote control, it autonomously finds its way and avoids pedestrians using vision alone. Its visual spatial intelligence navigation has approximately 120 million parameters and can support real-time on-device operation of up to 8 visual input channels at under 20 watts of power consumption. The team is adapting it for various types of robots, including delivery and pickup, inspection, agriculture, and cleaning.
Why read this
AI score 56
The original text provides the parameter count and power consumption metrics of the pure-vision on-device brain, allowing readers to gauge the threshold for deploying embodied navigation on the robot body itself.
Topics
Body · AI translation
At the 26th China International Industry Fair held in Shanghai, a white robot with "big eyes" and a chassis like that of a Mars rover caught people's attention. At the exhibition site, where crowds move about and scenes are complex, this robot has no pre-set map and does not rely on joystick remote control; using only its "eyes," it can find its own way, actively avoid pedestrians, and "tour the exhibition" together with visitors.
This robot is named "Danbao" (see photo, taken by Wei Pengfei). It is equipped with a pure-vision on-device brain and was jointly developed by the School of Future Information Innovation at Fudan University and the Jimu Tianshu team. Zhang Tianzi, head of the R&D team, introduced that for a long time, robot movement and operation have mainly relied on three methods—remote control, pre-programming, and pre-built mapping—rather than on the robot's own real-time understanding of the environment. Once the environment or task changes significantly, remapping or reprogramming is required. In operation scenarios such as emergency rescue and disaster relief and space mining, there are many changing factors, and robots that can understand the environment in real time and make autonomous decisions are urgently needed.
The birth of "Danbao" benefited from a breakthrough in pure-vision robot on-device brain technology. At the model level, the team unified various types of robot movement into a "navigation" problem: movement is navigating the body to a target position, while operations such as grasping are navigating the fingertips to near a target object. Many robots use two different systems for grasping and walking, which brings problems such as poor data consistency and relatively high cost, whereas "Danbao's" brain can command whole-body movement. Based on the above "on-device brain," a physical skill for embodied navigation—visual spatial intelligent navigation—has been trained and has already run stably on "Danbao."
Zhang Tianzi said that low computing power and high intelligence are another feature of visual spatial intelligent navigation. At present, embodied intelligent navigation large models mostly have parameter scales in the billions, which places high demands on computing power; some solutions rely on off-device inference and are difficult to deploy directly on existing robot bodies. Visual spatial intelligent navigation has about 120 million parameters and can support on-device real-time operation with up to 8 channels of visual input at a power consumption of less than 20 watts, freeing robots from dependence on pre-built maps and preset programs and moving them toward autonomous "understanding," decision-making, and action.
At present, the R&D team is adapting the results to various types of robots such as delivery and pickup, inspection, agriculture, and cleaning. In the future, it will explore enabling robots to continuously "evolve" from autonomous walking to autonomously completing tasks, and then to understanding intentions.