Non-parametric inference and coordination for distributed robotics

This paper presents non-parametric methods to infer the state of an environment by distributively controlling robots equipped with sensors. Each robot represents its belief of the environment state with a weighted sample set, which is used to draw likely observations to approximate the gradient of mutual information. The gradient leads to a novel distributed controller that continuously moves the robots to maximize the informativeness of the next joint observation, which is then used to update the weighted sample set via a sequential Bayesian filter. The incorporated non-parametric methods are able to robustly represent the environment state and robots' observations even when they are modeled as continuous-valued random variables having complicated multimodal distributions. In addition, a consensus-based algorithm allows for the distributed approximation of the joint measurement probabilities, where these approximations provably converge to the true probabilities even when the number of robots, the maximum in/out degree, and the network diameter are unknown. The approach is implemented for five quadrotor flying robots deployed over a large outdoor environment, and the results of two separate exploration tasks are discussed.