China News Service, Shanghai, October 12 (Reporter Chen Jing) On the 12th, at the 26th China International Industrial Fair (hereinafter referred to as: CIIF), the errand-running robot "Danbao" equipped with a pure visual robot end-side brain was officially released.
Zhang Tianzi, a postdoctoral fellow at the School of Future Information Innovation at Fudan University, said that breakthroughs in end-side brain technology for purely visual robots free robots from dependence on pre-built maps, etc., and will promote the underlying paradigm of robot autonomy from "reliance on presets" to "autonomous understanding." "Robots in the past acted according to presets or maps, but robots in the future will rely on intelligence to understand the world." Zhang Tianzi said.
For a long time, the movement and operation of robots have mainly relied on three methods: remote control, pre-programming and pre-mapping. Outside of remote control, whether it is a pre-set program trajectory or a pre-built map, it must be given before running, rather than coming from the robot's own real-time understanding of the environment. Once the environment or task changes, reprogramming or remapping is required.
It is understood that by relying solely on visual perception and completing spatial understanding and autonomous decision-making in real time on a terminal-side chip that consumes less than 20 watts of power, "Danbao" can navigate independently in complex and unfamiliar environments without the need for remote control or pre-built maps. The on-device brain allows the robot to directly understand the space it is in visually, and make decisions and actions autonomously based on this, without relying on real-time remote control, preset programs or pre-built maps.
According to reports, in terms of kernel design, the R&D team unified various movements and actions of the robot into "navigation" problems: movement is to navigate the body to the target location, and operations such as grabbing are to navigate the end effector to the vicinity of the target object. The first physics skill VSIN trained by the team based on the above-mentioned kernel has been implemented on the engineering prototype. Lightweight is an important feature of VSIN. "As the industry competes to improve the upper limit of robot capabilities, our team is committed to taking the lead in solving the lower limit of large-scale implementation of physical intelligence." Zhang Tianzi said. The parameter scale of current large-scale embodied intelligent navigation models is mostly in the billions, which requires high computing power. Some solutions require off-machine reasoning and are difficult to be directly deployed on existing robots. The number of VSIN parameters is about 120 million, which is less than one-tenth of similar large navigation models. It can support end-side real-time operation of up to 8 camera inputs with power consumption within 20 watts, providing a feasible path for large-scale deployment of robot physical intelligence.
It is also understood that the R&D team is adapting to various types of robots such as errand delivery, inspection, and commercial cleaning, and is adapting in parallel on domestic end-side computing chips. The team will follow the path of "space intelligence-task intelligence-end-side general intelligence", from letting the robot walk autonomously to completing tasks autonomously, to understanding intentions and continuing to evolve.
