Scaling Human-Robot Systems Prasanna Velagapudi Robotics Institute Carnegie Mellon University pkv+@cs.cmu.edu Paul Scerri Robotics Institute Carnegie Mellon University pscerri+@cs.cmu.edu INTRODUCTION Exciting applications are emerging that involve large, het- erogeneous human-robot teams acting in complex environ- ments. Examples include search and rescue [5], disaster re- sponse [12], and military applications [4]. Robots are capa- ble of augmentation and force multiplication of human as- sets, providing superhuman perception, coverage, and mo- bility, without risking human life. In these domains, the advantages of robotic teams lie pri- marily in their ability to cover large areas quickly, by co- ordinating and parallelizing their efforts. However, when equipped with the latest in multi-spectral imaging, 3D LI- DAR, and RF signal analysis, these platforms can also ac- cumulate vast amounts of information to process, reaching gigabits per second [11]. Available communication hard- ware is ill-equipped to handle such high data rates simul- taneously from many sources, and the problem is only ex- acerbated by power limitations and the overhead of wire- less network protocols, especially in remote and military set- tings [9]. In addition, in many of these domains, robots can be partially or sporadically connected, meaning that infor- mation must be buffered and relayed through other robots or other communications hubs. Overall, information from robots is bandwidth-constrained and latency can vary greatly. Under these conditions, the na¨ıve but common architecture of humans as overarching controllers quickly breaks down. Humans are a necessarily centralized resource, and it is im- practical and often impossible to simply pipe all necessary data for them to constantly control and correct autonomous robots. In addition, human attention and workload are tightly constrained resources, and the cognitive costs associated with rapid switching and real-time monitoring weigh heavily on operators, reducing their ability to effectively deal with many robots [7]. On the other hand, fully autonomous teams are deficient in several major areas, making human operators critical and likely to remain so in the near future. Human operators have Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. CHI 2009, April 3 - 9, 2009, Boston, MA, USA. Copyright 2009 ACM 978-1-60558-246-7/07/0004...$5.00. the advantage of a large database of meta-knowledge to aug- ment their reasoning skills. The ability to make use of this knowledge to perform inference is often necessary to inter- pret the data collected by robotic agents and convert it into useful hypotheses for the team. For example, in wilderness search and rescue, victims or signs of life can be extremely hard to spot directly, even by humans. Occlusions and un- structured terrain in the visual field limit visibility and add visual noise, e.g. Figure 1. However, humans are exception- ally good at spotting clues that are applicable in context. Pat- terns left by victims, potential obstacles, anomalous shapes or structures–the ability to notice these things as significant to the task is extremely difficult to codify, but trained humans excel at it. Figure 1. A wide range of unstructured terrain types complicates the interpretation of UAV video feeds. The high expense, but necessity, of human operators means that efficient human-robot teams must focus on letting these operators focus on the tasks that they do best, ones that au- tonomous systems cannot or should not do, while relegating the rest to robot autonomy. It is important to distinguish this from simply giving the autonomy components the “easiest” tasks: certain tasks, such as coordinated navigation, can be quite difficult for humans [16], but have many reasonable autonomous solutions [2]. Tasks relegated to humans face the additional cost of mov- ing information to and from human operators. Unlike au- tonomous software, humans cannot easily be moved between platforms or decentralized. The information they need to make their decisions must be shipped to them, and their de- 1