Active Robot Perception by Model-Based Policy Optimization with Safety and Visibility Constraints

In this talk, I will present techniques for planning and control of robots' sensing performance, known as active perception, while guaranteeing behavioral safety.

The first part of the talk will focus on active Simultaneous Localization and Mapping (SLAM), where a robot explores an unknown environment to construct an accurate and reliable map. I will present an adjoint method for open-loop planning, solved offline, that minimizes information-theoretic cost of the map by optimizing a time sequence of control inputs over continuous SE(3) space under a limited field of view. For closed-loop control executed online, a linear quadratic regulator (LQR) is applied to handle the robot's localization uncertainty by minimizing the covariance obtained from Extended Kalman Filter (EKF)-based SLAM.

The second part of the talk will cover policy learning and control methods for tracking moving targets. For scenarios without obstacles, I will demonstrate how to train a control policy for active target tracking using a model-based policy gradient over randomized configurations of multiple targets. This approach uses self-attention to parameterize the policy and a differentiable visibility objective to optimize it. For scenarios with obstacles, I will present a planning and control pipeline that maximizes the visibility of a moving target while addressing occlusion and safety requirements. This is achieved by introducing Control Barrier Functions (CBFs) for both visibility maintenance and collision avoidance.

Finally, I will share my future research vision for intelligent autonomous robots in more challenging scenarios, such as multi-agent active perception, dense mapping of unstructured environments, and cooperative autonomous driving.