Environment Model Adaptation for Mobile Robot Exploration

Many modern approaches to robotic exploration reason directly about the information that will be gained from future camera views. Although these approaches have been shown to be effective at reasoning about the effectiveness of exploration actions, evaluating information gain is computationally expensive. In this paper, we describe how to produce compressed representations of the environment that enable more efficient evaluation of information gain. We provide results in simulation and with a wheeled robot.

Read More
Efficient Online Multi-robot Exploration via Distributed Sequential Greedy Assignment

This paper describes a distributed planning approach for multi-robot information gathering and application to robotic exploration. Unlike popular sequential planners, the proposed approach gains in efficiency by allowing robots to plan in parallel. We prove that the proposed planner approaches the performance guarantee for sequential planning if we select good sets of plans from robots planning in parallel and demonstrate that we are able to do so in the simulation results.

Read More
Distributed Matroid-Constrained Submodular Maximization for Multi-Robot Exploration: Theory and Practice

This paper describes a distributed planning approach for multi-robot information gathering and application to robotic exploration. Unlike popular sequential planners, the proposed approach gains in efficiency by allowing robots to plan in parallel. We additionally extend this planning approach to consider the possibility of inter-robot collisions while avoiding deadlock and provide demonstrations in simulation and on a team of three quadrotor robots.

Read More
Active Estimation of Mass Properties for Safe Cooperative Lifting

This paper presents a method by which a team of aerial robots can lift an unknown object by learning about the mass distribution of the object while it is still on the ground. The robots are able to learn about the mass and center-of-mass of the object by applying forces at different locations and by tracking the force required to budge the object or else whether they are not able to do so. We then describe how to select highly informative interactions while simultaneously attaching robots in a configuration that is able to lift the object.

Read More
Distributed Submodular Maximization on Partition Matroids for Planning on Large Sensor Networks

This paper provides an efficient and distributed algorithm for planning for multi-robot sensor coverage. Sensor coverage is ubiquitous in robotics and models, for example, the objects that robots will observe with their cameras as they move through an environment. Popular sequential algorithms for multi-robot coverage planning guarantee good performance but require increasing numbers of planning steps for large teams of robots. We propose a randomized algorithm that uses only a fixed number of planning steps and approaches the performance of sequential planning when dependencies between agents decrease with distance.

Read More
Online planning for human – multi-robot interactive theatrical performance

This paper describes a full system for controlling multi-robot teams online, meaning that plans do not have to be designed prior to operation. The system is specifically illustrated with quadrotors in the context of an improvisational theatrical performance. The robot system interprets commands received from a human operator into dynamically feasible and collision free trajectories, and smoothly transitions from currently executing plans to the new trajectories in real time.

Read More
Interactive Online Choreography for a Multi-Quadrotor System

This short workshop paper gives an overview of our system enabling a theatrical performance, performed with six quadrotors over three acts, and dynamically directed by a theatric performer. This means that all behaviors performed by the robots during the course of the 100+ behavior production were not prescripted, but commanded, planned for, and executed online in real time during the production. Plots of the acceleration and minimum clearance distance over the robot team show that the generated trajectories are safe and dynamically feasible, and still images from the performance (accompanying video link provided) highlight elements of the approach.

Read More
Experience-driven Predictive Control with Robust Constraint Satisfaction under Time-Varying State Uncertainty

Vehicle dynamics and model parameters are rarely well known in practice. Robot controllers are further challenged by actuator constraints and noisy state estimates. This paper presents a controller that efficiently learns vehicle dynamics while at the same time efficiently ensuring that the vehicle satisfies state and input constraints in the presence of state uncertainty. Tracking error improvements are shown while maintaining constraint satisfaction and flying a quadrotor through a windy environment.

Read More
Choreographing Theatrical Scenes for Aerial Robots through Keyframe Interpolation

Artists and choreographers who wish to employ robots in theatrical productions may have trouble designing dynamically feasible robot trajectories. This short paper describes a simple design approach based on storyboarding, a common method for communicating spatial and temporal motions. Our approach allows a user to move physical robot tokens around a mock-up of the environment. A camera captures snapshots, or “keyframes,” of the user’s design, and a robotics expert or AI planner can then use well-understood trajectory generation methods to create robot trajectories from these keyframes.

Read More
RaD-VIO: Rangefinder-aided Downward Visual-Inertial Odometry

Aerial vehicles require a reliable means of determining how they move for robust flight. A downward looking camera is often used to provide an estimate of the robots motion along the ground. In this work we employ the approximation of the ground underneath as being locally planar and utilize this property in concert with onboard inertial sensors for measuring attitude and a single laser beam based rangefinder to provide fast, robust odometry estimates. Specifically, we include all the sensor information within a unified filtering framework. The focus in this work is to provide a reliable, fast, and robust estimate of odometry that can serve ably as a backup for more sophisticated optimization based odometry algorithms.

Read More
Reactive Collision Avoidance using Real-Time Local Gaussian Mixture Model Maps

A robot flying in unknown cluttered environments must be able to create a real-time map of its immediate surroundings and make sure that it does not collide into any obstacles. In this work, we propose representing the world as a mixture of Gaussian distributions, which enables us to represent the environment as a compressed and succinct representation. We use the geometric properties of this representation to enable efficient collision checking the robot surroundings.

Read More
Fast Monte-Carlo Localization on Aerial Vehicles using Approximate Continuous Belief Representations

A robot or a team of robots operating in large known environments require accurate knowledge of their exact location in the environment, in order to execute complex tasks. Limited size, weight and power constraints lead to constraints on the on-board computational capacity of aerial robots. This work presents a Monte-Carlo based real-time localization framework capable of running on such robots, enabled by using a compressed representation of the environment point cloud.

Read More
Persistent Robot Formation Flight via Online Substitution

This work presents a means of ensuring continuous operation by computing non-colliding paths to guide robots in and out of a planned, long duration path, allowing robots to temporarily depart and recharge. The focus of our approach is to quickly generate collision-free paths that minimally diverge from the original plan. In this manner, a user could generate an extremely long-duration plan that neglects battery capacity limitations and rely on our approach to swap depleted and charged robots, ensuring continuous operation.

Read More
Efficient Multi-Sensor Exploration Using Dependent Observations and Conditional Mutual Information

This paper presents a method to leverage conditionally dependent sensor observations for multi-modal exploration. A robot equipped with multiple sensors and potentially non-overlapping fields of view selects actions that drive the system to collect observations of occupied regions. The result is better coverage of the environment with the additional, dependent modality.

Read More
Computationally Efficient Information-Theoretic Exploration of Pits and Caves

This paper presents a method to enable exploration in significantly three-dimensional and complex subterranean environments such as caves. A simulated aerial system maximizes information-theoretic measures to maximize coverage of the environment. Information-theoretic measures are also used to compress the environment representation and enable faster exploration performance.

Read More