Learning for proximity flying – Proxilearn

The Proxilearn project leverages artificial intelligence techniques (machine learning) to allow micro-UAVs (10-20 cm / 50-80 g) to fly in very confined space (diameter between 40 cm and 1.5 m): air shafts, tunnels, underground storage, natural caves, or quarries. It is focused on two scientific challenges: (1) precise stabiliiazation in spite of the turbulences that result from the interaction between the rotors and the environment; (2) autonomous flight in low-light conditions with light, "minimalistic" sensors.

We will develop a stabilizer that exploits machine learning and data from a novel experimental setup to measure the force and and torques, based on a 7-DOFs manipulator. Using measures of forces and torques in many configurations (in a pipe, close to a wall, ...), we will train a probabilistic neural network that will predict the forces that are exerted on the quadrotor; then we will use this model to learn a control policy (a small neural network) that will be uploaded to the quadrotor.

In parallel, we will design an experimental quadrotor with a set of sensors for flying in confined environments, in particular an artificial retina and tiny laser-based distance sensors. We will combine these sensors with a novel algorithm to compute the 2-dimensional optic flow specially designed for this retina in low-light situations. We will then integrate all these sensors with an extended Kalman filter.

The third step will be to test and refine the autopilot in environments that are more and more narrow and less and less lit. The autopilot will rely on the stabilizer previously trained and on the states estimated with the extended Kalman filter. It will allow the UAV to explore, for instance, a pipe or an air shaft. The data recorded during the flight will be used to improve the force-torque model, which, in turn, will allow us to train a new, more effective policy.

On the long term, this kind of aerial robot will be capable of ignoring the complexity of the ground (steps, mud, water, ladder, ...) by moving in 3D and thus evolve in pipes, shafts, manholes, etc. The industrial impact is to be able to obtain images of hazardous sites; the impact for society is both to better take into account the risks raised by the underground (old undeground quarries, caves) and to propose an innovative tool for archaeology; the impact on defense and security is to increase the knowledge of a situation while using narrow openings and tunnels.