This volume is the second of two special journal issues compiled from the best papers presented at the fifth Robotics: Science and Systems (RSS) conference, held at the University of Southern California in June 2011. The first of these special issues was published by the International Journal of Robotics Research,

Since its inception in 2005, the conference has continued to attract scientists working on the algorithmic and mathematical foundations of robotics, robotics applications, and analysis of robotic systems. The papers presented in this special issue represent the best of robotics research in statistical inference, machine learning and planning, expanded and rigorously reviewed for journal publication. All papers share a common theme of providing new and fundamental algorithmic insights into the principles that govern how robots and intelligent machines operate in the real world.

1 Guide to the special issue

Over the years, RSS has been the venue for a number of major results in planning, and this year produced another set of advances. As we see robots needing to perform more complex tasks in human-oriented environments, manipulation and interaction with the environment has become an increasingly important area of study.

We begin with a paper by Dogar and Srinivasan who show that manipulation tasks can be performed by taking advantage of the dynamics of the objects. They address the problem of how a robot can manipulate objects by pushing, but the specific technical challenge is doing so in the presence of clutter and resulting uncertainty. Their work couples a physically realistic model of object dynamics with a set-based representation of the uncertainty, providing a robust model of the effects of pushing that can be incorporated into the grasp planner efficiently.

In a related vein, interacting with objects in the environment can often require highly dynamic control, especially for tasks such as catching and throwing. Such motions can be performed by varying the stiffness of the actuator, but the problem is how to control the actuator to achieve the desired task. Braun et al. provide a representation that allows the actuator profile to be optimised efficiently, directly exploiting the physical properties of the variable stiffness actuator. The result is not only a compelling demonstration for throwing, but also a path forward for complex explosive tasks.

A common assumption of many planning algorithms is that the underlying model is known completely; any changes in the environment or the vehicle usually necessitate generating a new plan from scratch. However, it is usually difficult if not impossible to justify the assumption of a known and perfect map. By planning with respect to uncertainty in the various models used for planning, the vehicle can not only behave robustly with respect to potential errors in the model, but can also take actions designed to improve the vehicle’s knowledge of itself and the environment. The difficulty is computing such plans efficiently, and we see in the paper by Kurniawati et al. an approach to motion planning with model uncertainty that directly addresses this problem using an alternate representation of the distribution over vehicle and environmental parameters, allowing a substantial reduction in the number of samples required to express possible future distributions. This reduction allows the planner not only to reason about much higher dimensional uncertainties than had been previously feasible, but also provides performance guarantees, opening the door to reliable autonomy under much more realistic assumptions of the prior knowledge of the world.

One of the long-standing limitations of motion planning is generating solutions quickly, and a popular approach is to reduce the planning problem to search on a graph that is a reduced representation of the environment. However, many paths through the graph are similar with respect to the coverage of the environment. When performing tasks such as exploration or search, ensuring that the search process generates a diversity of paths can be difficult, especially outside two dimensional domains. The work of Bhattacharya et al. provides a very general technique drawn by analogy from electromagnetism for generating paths that exist in different homotopy classes, allowing the search process to ensure that qualitatively different paths are explored, covering the environment efficiently.

Moving to a higher level of abstraction, we turn our attention to the problem of task allocation for multi-agent systems. The primary difficulty is again the computational complexity of solving assignment problems involving hundreds of robots and tasks. Liu and Shell observe that the partitioning techniques for balancing matrix computations across multiple processors can not only be applied to distribute the computation loads but, in fact, application of these techniques may actually be directly used to compute the solution to an assignment problem. The paper provides detailed results for task allocation for multi-agent systems of hundreds of robots in less than a second.

Continuing to consider high-level robot behaviour, the problem of guaranteeing the correctness of robot plans and their execution is starting to attract attention. While formal methods have been used to analyse general computation for some time, only in recent years have such techniques been applied to robotics. One of the primary difficulties in providing formal guarantees of correctness for robot behaviour has been the inherent uncertainty in sensor behaviour, and Johnson and Kress-Gazit give one of the first approaches to reasoning about robot behaviour that relax the requirement of perfect sensing. Their paper provides different approaches that, depending on the domain, allow either an exact computation of the probability of correctness, or provide upper and lower bounds on the probability.

Finally, Lindsey et al. close the special issue with a paper that truly advances the state of the art in quad-rotor operations, in which they describe a quad-rotor construction system that can build any instance of a class of structures known as special cubic structures. Their paper embodies the full scope of the conference, providing a theoretical analysis of the construction algorithm, the design of the system and experimental results of using quad-rotors to build complex structures.

Although the papers included in this volume were selected as representing RSS, all authors who presented papers at RSS 2011 are to be thanked for their support, and congratulated for representing the state of the art in robotics.