Drone shows are an increasingly popular form on the light display. These shows contain hundreds to thousands of bots in the air, each of which programmed flying on the paths, which together create complex shapes and patterns in the sky. When they go according to plan, drone shows can be spectacular. But when one or more drones work, just like recently in Florida, New York and elsewhere, they can pose a serious threat to viewers on earth.
Drone show accidents emphasize the challenges related to maintaining security in what engineers call “multi-aggregation systems”-many coordinated, cooperation and computer program, such as robots, drones and cars riding with their own drive.
Now the MIT engineers team has developed a training method for multi -aggregation systems that can guarantee their safe activity in crowded environments. Scientists have found that after applying the method for training a small number of agents, safety margins and controls learned by these agents can automatically scale to any larger number of agents, in a way that ensures system security as a whole.
During the demonstration in the real world, the team trained a small number of palm drones to safely perform various goals, from simultaneously switching position in half of power to landing on designated moving vehicles on Earth. In simulations, scientists have shown that the same programs, trained on several drones, can be copied and scaled to thousands of drones, enabling a large agent system to safely perform the same tasks.
“It can be a standard for any application that requires a team of agents, such as storage works, search drones and rescue and self -propelled cars,” says Fan Chuchu, Aeronautics Flight Professor and Astronautics in myth. “This provides a shield or safety filter, saying that every agent can continue his mission and we will tell you how to be safe.”
The fan and her colleagues report on their new method in the study appearing this month in the journal IEEE transactions regarding robotics. Co -authors of the study are students of MIT students Songyuan Zhang and Oswin, so there were Postdoc Kunal Garg myth, which is currently an adjunct for Arizona State University.
Margin MALL
When engineers design safety in any multi -aggregation system, they usually have to consider the potential paths of each agent in relation to any other agent in the system. This steam path planning is a time -consuming and expensive computing process. And even then safety is not guaranteed.
“In the drone show, each drone receives a specific trajectory – a set of points and a set of times – and then basically closes their eyes and follow the plan,” says Zhang, the main author of the study. “Because they only know where they have to be and what time, if there are unexpected things, they do not know how to adapt.”
Instead, the MIT team has developed the training method for a small number of agents for safe maneuvering, in a way that can effectively scale to any number of agents in the system. And instead of planning specific paths for individual agents, the method would allow agents to constantly map safety margins or boundaries outside which they can be dangerous. The agent can then take any number of paths to perform his task, as long as he stays within his safety margins.
In a sense, the band claims that the method is similar to how people intuitively move in the environment.
“Say you are in a really crowded shopping center,” he explains. “You don't care about anyone except people who are in your immediate vicinity, like 5 meters surrounding you, in terms of safe movement and not falling into anyone. Our work has a similar local approach.”
Safety barrier
In its new study, the team presents their method, GCBF+, which means “chart control barrier function”. The barrier function is a mathematical term used in robotics, which calculates the type of safety barrier or border, after which the agent is likely to be dangerous. In the case of any agent, a security zone can change the moment for a moment, because the agent moves among other agents who move in the system themselves.
When designers calculate the barrier functions for any agent in a multi -stage system, they usually have to take into account potential paths and interactions with any other agent in the system. Instead, the MIT MIT method calculates the security zones only a handful of agents in a way that is accurate enough to represent the dynamics of many other agents in the system.
“Then we can, in a sense, copy this barrier function for each agent, and then suddenly we have a chart of security zones that work for any number of agents in the system,” he says.
To calculate the agent's barrier function, the team method first takes into account the “radius of detection” of the agent or how much the environment can be observed by the agent, depending on the sensor's capabilities. As in the analogy of the shopping center, scientists assume that the agent cares about agents who are within the detection radius in terms of safety and avoiding collision with these agents.
Then, using computer models that capture the special functions and limits of the agent, the team simulates the “controller” or a set of instructions on how an agent and a handful of similar agents should move. Then they simulate many agents moving on some trajectories and register whether and how they collide or otherwise interact.
“When we have these trajectories, we can calculate some of the laws that we want to minimize as we say how many security violations we have in the current controller,” says Zhang. “Then we update the controller to make it safer.”
In this way, the controller can be programmed to real agents, which would allow them to constantly map the security zone based on other agents that they can feel in their immediate surroundings, and then move in this security zone to perform their task.
“Our controller is reactive,” says the fan. “We do not plan the path before. Our controller constantly takes information about where the agent is going, what his speed is, how quickly other drones go. He uses all this information to develop a plan in flight and each time replacing. So, if the situation changes, he is always able to adapt to safety.”
The team demonstrated GCBF+ in the system of eight crazy flowers-four-headed drones of the size of the hands, whose task is to fly and switch position in the air. If the drones were to do this, following the simplest path, they would certainly collide. But after training with the band's method, they were able to introduce around them in real time to maneuver around them, keeping their safety zones to effectively change positions in flight.
In a similar way, the team ordered the drone flying, and then landing on specific turtles-wheeled turtles with tops reminiscent of a shell. Turtlebots still rode in a large circle, and crazy flowers could avoid collision with each other when they made a landing.
“Using our frames, we only need to give drones our destinations instead of the entire collision -free trajectory, and drones can learn how to reach their places without a collision,” says a fan, which predicts that the method can be applied to any system of multi -person delivery systems.
These works were partly supported by the American National Science Foundation, myth Lincoln Laboratory as part of the Aerobatic Flights (SAFR) security program and the Defense Agency and the Technology Agency in Singapore.
