A recent paper by the members of the DCIST alliance develops a method for continuous-space optimal control of active information acquisition. They have developed “iterative Covariance Regulation (iCR)”, a novel method for an information-theoretic active perception performing multi-step forward-backward gradient descent. The problem is formalized as SE(3) trajectory optimization over a multi-step continuous control sequence of a robot’s linear and angular velocity inputs to minimize the differential entropy of a map state conditioned on a sequence of measurements (e.g., Lidar or RGB-D camera). To ensure that the covariance matrix evolution is differentiable with respect to the control sequence, they introduced a new differentiable field of view formulation for the sensing model, providing a smooth transition from unobserved to observed space in the environment. Finally, the gradient of the objective function with respect to the multi-step control input sequence is computed explicitly and the control trajectory is updated via gradient descent. iCR algorithm was tested in simulated active mapping experiments in comparison with two baseline methods and they observed that iCR achieves significantly larger reduction of the map uncertainty due to its continuous-space optimization.

Capability: T3C1D: Optimal control and reinforcement learning with information theoretic objectives

Points of Contact: Nikolay Atanasov (PI), Shumon Koga, and Arash Asgharivaskasi

Video: https://youtu.be/Qgt64a0bOiA

Paper: https://arxiv.org/pdf/2103.05819.pdf

Citation: S. Koga, A. Asgharivaskasi, and N. Atanasov “Active Exploration and Mapping via Iterative Covariance Regulation over Continuous SE(3) Trajectories”, In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2021.