In recent years, the fast navigation of autonomous vehicles has received increasing interest. For several years, autonomous drone racing competitions have also been organized in which thousands of people around the world participate. The state-of-the-art algorithms in estimation and control are currently reaching a very high level of maturity. This involves the need for trajectory planning algorithms that take full advantage of the vehicle's capabilities while taking into account the limitations of vehicle dynamics, rather than relying on simplified models. A drone has been developed in MIT that avoids obstacles efficiently even at high speeds.
Study of the feasibility constraints at MIT for the drone
MIT scholars have developed an algorithm capable of obtaining a dynamically feasible and optimal quadrotor trajectory over time and able to help drones avoid obstacles at high speeds to avoid collisions . The team of researchers believes that the algorithm can be used in operations that require very short times in more complex situations such as, for example, the search for survivors of a natural disaster.
Therefore, precise knowledge of dynamic feasibility constraints is required to find the optimal time. This complicates the problem, as these feasibility constraints can become highly complex when flying at high speed.
The maneuvers are influenced by the dynamics of the flight , but also by hardware and software for controlling and estimating the state of the system . The resulting set of feasibility constraints with memory (i.e. which also take into account previously made flights) cannot be easily incorporated into a typical flight planner for two main reasons.
First, feasibility constraints are not easily expressed in a convenient way, for example, as constraints on an admissible set of inputs and control states. Instead, the feasibility of the trajectory must be considered holistically. Second, in most scenarios, precise modeling of these constraints is only possible through real-world experiments.
These experiments are risky and potentially costly, since the process, aimed at improving the understanding of how high-speed aerodynamics can affect drones in flight, involves numerous laboratory experiments, in which collisions and therefore complete or partial destruction often occur. of drones . These experiments in fact turn out to be particularly expensive. For this reason, the MIT team's algorithm aims to minimize the number of experiments required to identify safe and fast routes for a drone.
The novelty of the algorithm
Many trajectory optimization schemes rely on a large number of evaluations. The algorithm created by the MIT engineers, on the other hand, uses a multi-fidelity optimization technique capable of approximating the feasibility constraints of the system based on a limited number of experiments. It uses a black-box model of the Gaussian process (GP) to classify candidate trajectories as feasible or not feasible and is therefore able to plan faster and faster trajectories as the model improves.
The MIT scholars therefore proposed for modeling the feasibility constraints of the quadrotor and for the generation of trajectories with optimal timing. All this was possible thanks to the use of a multi-fidelity GP classification algorithm that can incorporate evaluations from analytical approximation, numerical simulation and flight experiments in the real world by designing an acquisition process specifically adapted to experimental robotics. The acquisition function takes into account the additional cost of unachievable assessments, as these can pose a threat to the vehicle and its surroundings.
The experiments carried out at MIT to train the drone
The first experiments conducted were able to simulate the behavior of a drone in a flight in the presence of virtual obstacles. Thousands of different scenarios with different routes and flight speeds were then simulated.
To demonstrate the approach, the team said they simulated a drone flying through a simple path with five square obstacles offset from each other. They set up the same setup in a physical training space and programmed a drone to fly through the course at speeds and trajectories identified by their simulations. They also ran the same route with a drone trained on a more conventional algorithm that doesn't incorporate experiments into its planning.
The drone trained on the new algorithm “won” every race , completing the course faster than the conventionally trained drone, in some cases, finishing the course 20% faster than its competitor despite a slower start.
The researchers plan to fly more experiments, at higher speeds and through more complex environments, to further improve the algorithm . They will also be able to incorporate flight data from human pilots driving drones remotely, whose decisions and maneuvers could help focus on faster but still feasible flight plans.
Article by Giorgia Pascale