International audience

In this paper, a Deep Reinforcement Learning (DRL) algorithm is proposed, to learn a control policy for dynamic systems with input and state constraints. The system dynamics are assumed to be unknown for control design, the only available a priori information being the formulation of input and state constraints. Elements of Control Theory are leveraged to obtain safety certificates along with the learned policy: Control Lyapunov Functions (CLF) for closed loop stability, and Control Barrier Functions (CBF) for state constraint satisfaction. These two notions are transformed into conditions to be numerically verified in a PPO-based DRL algorithm, where a neural network CLF is also learned along with the control policy. Simulation results are presented for two application examples, illustrating closed-loop stability and constraint satisfaction resulting from the learned controllers.

This article considers the cooperative search and track of moving targets by unmanned aerial vehicles (UAVs) over some regions of interest (RoIs). The RoI contains obstacles that may partly hide the targets. UAVs are not aware of the obstacles' locations and cannot determine when an obstacle limits their field of view. In this article, no map of the RoI is built. This reduces the computational complexity but makes the selection of an appropriate point of view to observe a specific part of the RoI difficult. Showing the absence of a target at a given location of the RoI is then challenging. To address this problem, we introduce a detectability set for each point of the RoI as the set of all UAV locations from where that point is visible. The detectability sets are unknown to the UAVs and evolve with time when the environment is time varying. We assume that each detectability set contains at least one half-cone with a minimal aperture, which translates the fact that targets are never fully occluded by the environment. We prove that a finite number of$L$simultaneous observations are sufficient to guarantee the presence or absence of targets at a given location. This leads to a partition of the fleet into groups of$L$UAVs that simultaneously collect measurements on a specific zone of the RoI. Assuming that state perturbations and measurement noises are bounded, a distributed set-membership estimator is used to evaluate set estimates for potential target locations. The trajectories of the UAVs are designed using a model-predictive control approach to reduce the estimation uncertainty. Simulations illustrate the performance of the proposed approach in the presence of unknown static or moving obstacles.

This paper presents experimental results of cooperative guidance laws embedded on Lego Mindstorms NXT mobile robots for two types of missions. The first one is navigation to a waypoint as a fleet with collision and obstacle avoidance, following a model predictive control (MPC) framework. The second one is source localization, i.e., finding the maximum of a potential field, for which a distributed estimation and control strategy is proposed. Experiments show the ability to perform the two missions on these basic mobile robots, in spite of their limited computational resources. In particular, the search for the optimal control sequence through a dedicated discretization of the command space makes it possible to implement real-time MPC.

This paper focuses on viewpoint planning for 3D active object recognition. The objective is to design a planning policy into a Q-learning framework with a limited number of samples. Most existing stochastic techniques are therefore inapplicable. We propose to use Kriging and bayesian Optimization coupled with Q-learning to obtain a computationally-efficient viewpoint-planning design, under a restrictive sampling budget. Experimental results on a representative database, including a comparison with classical approaches, show promising results for this strategy.

This paper presents a method to design guidance and control laws for small vertical take off and landing unmanned aerial vehicles when no measurement of linear velocity or angular velocity is available. The control strategy is based on the introduction of virtual states in the state equation of the system and allows the design of stabilizing feedback controllers without using any observer. Simulation results are provided for six degrees of freedom model of a small rotorcraft-based unmanned aerial vehicle.