Autonomous Robots 8,1-6 (2001)  2001 Kluwer Academic Publishers. Manufactured in The Netherlands. * The research described in this paper was carried out at the Jet Propulsion Laboratory, California Institute of Technology, under a contract with the National Aeronautics and Space Administration. Biologically Inspired Autonomous Rover Control* TERRY HUNTSBERGER Jet Propulsion Laboratory, MS 82-105, 4800 Oak Grove Drive, Pasadena, CA 91109 [email protected] Abstract. Robotic missions beyond 2013 will likely be precursors to a manned habitat deployment on Mars. Such missions require robust control systems for long duration activities. Current single rover missions will evolve into deployment of multiple, heterogeneous cooperating robotic colonies. This paper describes the map-making memory and action selection mechanism of BISMARC (Biologically Inspired System for Map-based Autonomous Rover Control) that is currently under development at the Jet Propulsion Laboratory in Pasadena, CA [Huntsberger and Rose, 1998]. BISMARC is an integrated control system for long duration missions involving robots performing cooperative tasks. Keywords: biologically inspired control, action selection mechanisms, mobile robots 1. Introduction Robotic outposts and precursor missions for deployment of manned habitat infrastructure on Mars are being studied by NASA for the second decade of this century [Schenker, et al., 2000; Huntsberger, et al., 2001]. Such missions require more autonomy in their control architecture than the Pathfinder/Sojourner mission that landed on Mars in the summer of 1997. Representative biologically inspired navigation and control systems for such outposts include CEBOT [Fukuda and Kawauchi, 1993], Q-machines [Kube and Zhang, 1997], the Tropism System Cognitive Architecture [Agah and Bekey, 1997], behavior-based control [Mataric, 1997], ALLIANCE [Parker, 1998], and CAMPOUT [Pirjanian, et al., 2000 and references therein]. Good overviews can be found in [Maes, 1991; Pfeifer and Scheier, 1999]. We have recently developed a multi-robot control architecture called BISMARC (Biologically Inspired System for Map-based Autonomous Rover Control) for long duration missions [Huntsberger and Rose, 1998]. It is based on a free-flow hierarchy (FFH) similar to the DAMN architecture [Rosenblatt and Payton, 1989], and has been used successfully for a number of different simulated mission scenarios including multiple cache retrieval [Huntsberger, 1997], fault tolerance for long duration missions [Huntsberger, 1998], and site preparation [Huntsberger, et al., 1999]. The system includes all aspects of safety, self-maintenance, and goal achievement that robotic systems require for a sustained planetary surface presence. It currently doesn’t include global planning or any adaptive learning capabilities beyond map-making. The next section briefly describes the organization of BISMARC, followed by a discussion of the action selection mechanism2 Huntsberger and map-making memory of the system. We close with experimental studies and conclusions. 2. BISMARC Organization BISMARC is organized as a two level system (shown in Figure 1). The first level generates possible motor actions using stereo images and the second level uses these action hypotheses coupled with external and internal inputs to drive the actuators on the robot. The DriveMaps algorithm used for action generation analyzes local range information for clear paths relative to a goal and is currently implemented on the SRR and FIDO rovers at JPL. A fuzzy adaptive behavior system with similar capabilities to DriveMaps is described in [Tunstel, 2001]. Figure 2 illustrates an action selection hierarchy for a cache retrieval mission. The rectangular boxes represent behaviors and the ovals are sensory inputs (either fixed, direct, or derived). At the top are the high level behaviors including Avoid Dangerous Places, Sleep at Night, Warm Up, Scan for Cache, Get Cache, Figure 1. Two level BISMARC architecture with stereo processing, action generation, and action selection subsystems. Figure 2. Free-flow hierarchy action selection mechanism for cache retrieval mission scenario. Ovals represent inputs derived from sensory stimuli, rectangular boxes are behaviors, and circles are temporal and uncertainty penalties. All weights on inputs to behaviors are 1.0 unless otherwise noted. Segmented boxes and ovals represent directional inputs (only cardinal directions shown but in practice continuous coverage). See text for further details.Biologically Inspired Autonomous Rover Control 3 Cool Down, Get Power, and Keep Variance Low. These goals are related to both task and rover safety. For example, since most planetary surface rovers have only visual sensors for navigation, the sensory input for Proximity to Night is derived from knowledge of the sun’s position and forces the rover to sleep at night by weighting the input to Sleep at Night heavier (16.0) than any other behavior in the hierarchy. The intermediate level behaviors are designed to interact with both the short term memory (STM), which corresponds to perceived sensory stimuli, and the long term memory (LTM), which encodes remembered sensory information. Control loops are prevented through temporal penalties (shown as T-circles in Figure 2) that constrain the system to only repeat a behavior a predetermined number of times. The bottom level behaviors in the hierarchy fuse the sensory inputs and the activations of the higher level behaviors in order to select an appropriate action to drive the actuators. The next section describes the action selection mechanism of BISMARC in more detail. 3. Action Selection Mechanism Action selection mechanisms for rovers on a planetary surface require low computational overhead, reactivity even in uncertain environments, no loss of internal state information, combination of conflicting behaviors, and the localization of sensory input to the appropriate modules. The FFH used in BISMARC includes all of these capabilities. Combining the inputs to a behavioral node is usually calculated as a simple weighted summation. This approach leads to potential problems in the case where the same goal triggers two or more behaviors and the utility of a behavior lower in the hierarchy should not be the sum of their activations. For example, in Figure 2, the Get Cache goal feeds into the Approach Perceived Cache and the