We propose a decentralized variant of Monte Carlo tree search (MCTS) that is suitable for a variety of tasks in multi-robot active perception. Our algorithm allows each robot to optimize its own actions by maintaining a probability distribution over plans in the joint-action space. Robots periodically communicate a compressed form of their search trees, which are used to update the joint distribution using a distributed optimization approach inspired by variational methods. Our method admits any objective function defined over robot action sequences, assumes intermittent communication, is anytime, and is suitable for online replanning. Our algorithm features a new MCTS tree expansion policy that is designed for our planning scenario. We extend the theoretical analysis of standard MCTS to provide guarantees for convergence rates to the optimal payoff sequence. We evaluate the performance of our method for generalized team orienteering and online active object recognition using real data, and show that it compares favorably to centralized MCTS even with severely degraded communication. These examples demonstrate the suitability of our algorithm for real-world active perception with multiple robots.

We present an active simultaneous localization and mapping~(SLAM) framework for a mobile robot to obtain a collision-free trajectory with good performance in SLAM uncertainty reduction and in an area coverage task. Based on a model predictive control~(MPC) framework, these two tasks are solved by the introduction of a control switching mechanism. For SLAM uncertainty reduction, graph topology is used to approximate the original problem as a constrained non-linear least-squares problem. A convex half-space representation is applied to relax non-convex spatial constraints that represent obstacle avoidance. Using convex relaxation, the problem is solved by a convex optimization method and a rounding procedure based on singular value decomposition (SVD). The area coverage task is addressed with a sequential quadratic programming (SQP) method. A submap joining approach, called Linear SLAM, is used to address the associated challenges of avoiding local minima, minimizing control switching, and potentially high computational cost. Finally, various simulations and experiments using an aerial robot are presented that verify the effectiveness of the proposed method, showing that our method produces a more accurate SLAM result and is more computationally efficient compared with multiple existing methods.

Understanding animal movements that underpin ecosystem processes is fundamental to ecology. Recent advances in animal tags have increased the ability to remotely locate larger species; however, this technology is not suitable for up to 70% of the world’s bird and mammal species. The most widespread technique for tracking small animals is to manually locate low-power radio transmitters from the ground with handheld equipment. Despite this labor-intensive technique being used for decades, efforts to reduce or automate this process have had limited success. Here, we present an approach for tracking small radio-tagged animals by using an autonomous and lightweight aerial robot. We present experimental results where we used the robot to locate critically endangered swift parrots (Lathamus discolor) within their winter range. The system combines a miniaturized sensor with newly developed estimation algorithms to yield unambiguous bearing- and range-based measurements with associated measures of uncertainty. We incorporated these measurements into Bayesian data fusion and information-based planning algorithms to control the position of the robot as it collected data. We report estimated positions that lie within about 50 meters of the true positions of the birds on average, which are sufficiently accurate for recapture or observation. Further, in comparison with experienced human trackers from locations where the signal was detectable, the robot produced a correct estimate as fast or faster than the human. These results provide validation of robotic systems for wildlife radio telemetry and suggest a way for widespread use as human-assistive or autonomous devices.

This paper presents an extension of Gaussian process implicit surfaces (GPIS) by the introduction of geometric object priors. The proposed method enhances the probabilistic reconstruction of objects from three-dimensional (3-D) pointcloud data, providing a rigorous way of incorporating prior knowledge about objects expected in a scene. The key ideas, including the systematic use of surface normal information, are illustrated with one-dimensional and two-dimensional examples, and then applied to simulated and real pointcloud data for 3-D objects. The performance of our method is demonstrated in two different application scenarios, using complete and partial surface observations. Qualitative and quantitative analysis of the results reveals the superiority of the proposed approach over existing GPIS configurations that do not exploit prior knowledge.