梦晨 发自 凹非寺
这项来自浙江大学的研究成果登上最新一期Science Robotics封面。
据浙江大学介绍,此前的机器人集群表演大多通过卫星定位和轨迹编码实现,由地面计算机统一控制。
这种模式下,机器人群体一旦失去指挥就会“群龙无首”,不但无法保持队形还可能撞上障碍物或相互碰撞。
这次的新成果被Science Robtics评价为第一个能在非结构化环境中分散、自主飞行的集群系统。
可以在避障之后迅速恢复队形。
也可以相互配合持续追踪特定目标。
那么浙大团队是怎么做的呢?
*(完整演示视频在文末)*
鸟群模式
论文中介绍,飞行机器人的研究受动物启发,分为虫群模式和鸟群模式两种。
昆虫会做短程的反应性动作(比如苍蝇避开苍蝇拍)。
基于反应的虫群导航算法对算力和内存的需求更小,机器人可以做得更小。
鸟类有更敏锐的感官和更大的脑容量,可以做长期的轨迹规划。
基于轨迹规划的鸟群导航算法有更强的性能和可扩展性,因此浙大团队选择了这一种。
在群体轨迹规划算法上,如果只考虑空间因素会影响机器人集群间的配合。比如通过狭窄空间时会拥堵,导致后面的机器人必须绕路。
因此,浙大团队同时对时间和空间做轨迹规划,利用稀疏参数优化(sparse parametric optimization)和约束转录(constraint transcription)方法提升速度,做到实时计算。
在穿越高密度竹林时,这种算法可以让多个机器人先后通过狭窄缝隙避免碰撞,无惧倾斜竹子和高低起伏的地形。
除了轨迹规划外,浙大团队改进了视觉-惯性里程计 (Visual-Inertial Odometry)做群体的定位。
为了避免长距离积累的微小误差最终造成相互碰撞,开发了分布式漂移校正算法。
每个机器人都有完整的感知、定位、规划和控制功能,相互之间用高保真无线通信来共享轨迹。
在10个机器人密集飞行实验中,研究人员关闭了GPS信号、临时增加障碍物、以及人类主动干扰都没出现碰撞。
实验所用的机器人由浙江大学控制科学与工程学院和湖州实验室研发。
单个机器人只有手掌大小,比一听可乐的重量还要轻。
搭载了英伟达Xavier NX模组,拥有6核CPU和384核GPU及8GB内存。
但在实验中,除了个别例外场景,CPU和GPU的使用率都保持在40%以下,在有限的计算资源实现了复杂行为。
将用于救灾、勘探和运输
论文第一作者为浙江大学控制科学与工程学院博士研究生周鑫,通讯作者为该院高飞博士和许超教授。
团队成员来自科学与工程学院及湖州研究院。
这次成果解决了在混乱的野外环境中机器人集群自主导航的问题,提高了对各种现实任务的适应性。
地震、洪水和火灾中,机器人集群可用于搜索、引导受困者,或运送紧急物资。
在生态研究、地质勘探中,使用机器人集群可以调查狭窄的环境。
而开发出的自主导航算法,也可以用于火星车、月球车,以及多台货运无人机协作运输重量超过单台运输能力的货物。
Mengchen comes from au Fei Temple
The research from Zhejiang University is on the cover of the latest issue of Science Robotics.
According to Zhejiang University, previous robot cluster performances are mostly realized by satellite positioning and trajectory coding, and are uniformly controlled by ground computers.
In this mode, once the robot group loses command, it will be “leaderless”. It will not only be unable to maintain formation, but may also hit obstacles or collide with each other.
This new achievement is evaluated by Science Robtics as the first cluster system that can fly independently in an unstructured environment.
You can quickly return to formation after avoiding obstacles.
You can also work with each other to track specific targets on a continuous basis.
So how did the team of Zhejiang University do it?
* (full demo video at the end of the article) *
Bird flock model
In this paper, the research of flying robot is inspired by animals, which can be divided into two kinds: insect swarm mode and bird swarm mode.
Insects can do short distances.Reactive action(for example, flies avoid fly swatters).
Based on inverseShouldInsect swarm navigation algorithm based on webThere is less need for computing power and memory, and robots can do it even smaller.
Birds have sharper senses and larger brain capacity and can do long-termTrajectory planning.
Bird flock navigation algorithm based on trajectory planningIt has stronger performance and scalability, so the team of Zhejiang University chose this one.
In the group trajectory planning algorithm, if only space factors are considered, the cooperation between robot clusters will be affected. For example, there will be congestion when passing through the narrow space, causing the robot behind to make a detour.
Therefore, the team of Zhejiang University makes trajectory planning for time and space at the same time, and uses sparse parameter optimization (sparse parametric optimization) and constrained transcription (constraint transcription) methods to improve speed and achieve real-time calculation.
When walking through high-density bamboo forests, this algorithm allows multiple robots to avoid collisions through narrow gaps, without fear of tilting bamboo and undulating terrain.
In addition to trajectory planning, the team of Zhejiang University has improvedVision-inertia odometer (Visual-Inertial Odometry) do group positioning.
In order to avoid collision caused by small errors accumulated over a long distance, a distributed drift correction algorithm is developed.
Each robot has complete sensing, positioning, planning and control functions, and uses high-fidelity wireless communication to share tracks with each other.
In 10 dense robot flight experiments, the researchers turned off the GPS signal, temporarily increased obstacles, and human active interference did not collide.
The robot used in the experiment was developed by the School of Control Science and Engineering of Zhejiang University and Huzhou Laboratory.
A single robot is the size of a palm, which is lighter than a can of cola.
Equipped with Nvidia Xavier NX module, with 6-core CPU and 384-core GPU and 8GB memory.
However, in the experiment, with the exception of a few exceptional scenarios, the utilization rate of CPU and GPU remains below 40%, and the complex behavior is realized in the limited computing resources.
Will be used for disaster relief, exploration and transportation
The first author is a doctoral student from the School of Control Science and Engineering, Zhejiang University.Zhou XinThe newsletter is written by the hospitalGoofyThe doctor andXu ChaoProfessor.
The team members are from the College of Science and Engineering and Huzhou Research Institute.
This result solves the problem of autonomous navigation of robot clusters in the chaotic field environment, and improves the adaptability to various practical tasks.
In earthquakes, floods and fires, robot clusters can be used to search for, guide people in distress, or transport emergency supplies.
In ecological research and geological exploration, robot clusters can be used to investigate narrow environments.
The autonomous navigation algorithm developed can also be used for rovers, lunar rovers, and multiple freight drones to cooperate to transport goods that exceed the capacity of a single vehicle.