NEUROBIOLOGICALLY-BASED CONTROL SYSTEM FOR AN ADAPTIVELY WALKING HEXAPOD* WILLIAM LEWINGER EECS Department, Case Western Reserve University, 10900 Euclid Ave Cleveland, OH, 44106, USA ROGER QUINN EMAE Department, Case Western Reserve University, 10900 Euclid Ave Cleveland, OH, 44106, USA Biological systems such as insects have often been used as a source of inspiration when developing walking robots. Insects’ ability to nimbly navigate uneven terrain, and their observed behavioral complexity have been a beacon for engineers who have used behavioral data and hypothesized control systems to develop some remarkably agile robots. However, it is now possible to implement models of relatively recent discoveries of the stick insect’s local control system (its thoracic ganglia) for hexapod robot controllers. Walking control based on a model of the stick insect’s thoracic ganglia, and not just observed insect behavior, has now been implemented in a complete hexapod able to walk, perform goal-seeking behavior, and obstacle surmounting behavior, such as searching and elevator reflexes. Descending modulation of leg controllers is also incorporated via a head module that modifies leg controller parameters to accomplish turning in a role similar to the insect’s brain and subesophageal ganglion. While many of these features have been previously demonstrated in robotic subsystems, such as singleand two-legged test platforms, this is the first time that the neurobiological methods of control have been implemented in a complete, autonomous walking hexapod. This paper discusses the implementation of the biologically-grounded insect control methods and descending modulation of those methods, and demonstrates the performance of the robot for navigating obstacles and performing phototaxis.

1. Introduction Because of their extreme mobility and agile adaptability to irregular terrain, insects have long been an inspiration for the designers of mobile, legged robots. Early hexapod robots such as Genghis [3] and later creations such as Tarry [7] implemented insect-like mobility based on observations of insect behaviors. The inter-leg coordination system developed by Holk Cruse [4], [5] has been *

This work is supported in part by NSF IGERT Training Grant DGE 9972747 and Eglin AFB Grant FA9550-07-1-0149. 1

2

widely implemented in legged hexapods and its basis is in behavioral experiments that qualitatively analyzed insect walking behaviors. These robots are impressive in their ability to navigate irregular terrain, track goal-based objects, and mimic insect-like walking. However, their functionality is based on observed insect walking, and not the underlying neurobiological control system. Neurobiological control methods have previously been demonstrated for simulated hexapods [6], [10] and mechanical subsystems, such as single- and two-legged test platforms [9], but not yet as a complete, mechanical hexapod. The hexapod described in this paper, BILL-Ant-a (Biologically-Inspired Legged Locomotion – Ant – autonomous) (Fig. 1), is the first physical, autonomous hexapod that uses a distributed walking control system based on the neurobiology of stick insects. A modeled biological network coordinates the joints in each leg and generates the desired stepping actions based on internal pattern generators that are modified by external influences from the environment. By employing environmental cues to control stepping behaviors, the robot is fully adaptive to irregular terrain and is also able to navigate obstacles using insect-based reflexes such as the elevator and searching reflexes.

Figure 1. BILL-Ant-a (Biologically-Inspired Legged Locomotion – Ant – autonomous), an autonomous hexapod with a neurobiologically-based control system for adaptive walking.

An additional module performs some of the functions of the insect’s brain and subesophageal ganglion. This module monitors and actuates a pair of light-

3

sensitive receptors in response to an illuminated goal. When the goal is not located before the robot, a neck actuator rotates to locate the direction of the goal. Then, the same module uses descending modulation to uniquely alter the parameters of the six leg controllers. These modified parameters allow the robot to perform turning maneuvers, much like an actual insect. 2. Control of Walking 2.1. Intra-Leg Joint Coordination As has been observed in the neurobiology of stick insects, each joint of a multijointed leg is controlled by its own pattern generator. These pattern generators cyclically oscillate individual joints at a natural frequency. By coupling sensory information associated with a joint, and/or sensory information associated with the other joints in the leg, the pattern generators can be influenced in order to create smooth, synchronized stepping movements. These neurobiological studies have been described in [8], modeled in [6] and have been condensed into a walking control method for 3-DOF legs [9]. Sensory feedback from leg load and joint angles coordinate the joints and generate stepping motions. Environmental cues, such as sensing ground contact on irregular terrain, cause the stepping pattern to dynamically alter itself, without the need for either complex calculations or high-level micromanagement of joint actions, thus producing a robustly adaptive robot. 2.2. Inter-Leg Gait Generation Another consideration for insects and walking robots is the coordination of stepping actions to produce stable, yet adaptive gaits. Work done throughout the years by Holk Cruse [4], [5] has identified mechanisms that may be responsible for coordinating the switching between swing and stance phases such that smooth, speed-dependent transitions between different gaits can emerge (such as migrating from a wave gait to a tetrapod gait to an alternating tripod gait). The implementation of this method, however, can be computationally intensive. A new method based on Cruse’s work has been developed [10] that simplifies gait generation, while maintaining the key factor of speed-dependent transitions between gait patterns. Although this new method is not as extensive in its capabilities as Cruse’s method, it is nonetheless sufficient for generating dependable, stable gaits for a legged hexapod.

4

2.3. Implementation Both intra-leg joint coordination and inter-leg gait generation are implemented on BILL-Ant-a through a network of distributed, low-computation-capable microcontrollers; each microcontroller is responsible for a single leg, akin to the function of thoracic ganglion segments in insects. While there are many microcontrollers currently available that provide impressive computing abilities, the low-end microcontrollers used by this robot were specifically chosen to demonstrate the computational simplicity of this leg control method. Each leg controller cyclically performs three sets of actions: gathering leg data (joint load and joint angle information); deciding in what direction each joint will be moving based on the coupled sensory information from self-, interjoint, and environmental-influencing signals; and then actuating each joint according to base-level muscle activation values (responsible for determining the maximum speed at which a joint can move) and a simple, piece-wise linear muscle model (used to create smoother walking movements). During the data gathering phase, legs share stance/swing state and thoraxcoxa joint (used to protract and retract the leg) angle information with orthogonal neighbors. These data are then used to influence the stepping cycles of neighboring legs in order to create stable gaits. Gait coordination relies on two simple rules. The first rule uses the stance/swing state to prevent legs from entering swing phase while at least one neighbor is already in swing. By delaying the stance/swing transition, neighboring legs are prevented from being in swing simultaneously, which maintains static stability. The second rule employs a ramp function generated by the thorax-coxa joint angle. This is used to determine how far along a nominal step length a leg has moved. If both the current and neighbor legs are approaching their respective stance/swing transition points and the neighboring leg slightly lags in the completion of its stance phase, the current leg will step early. This action minimizes, or eliminates, the need for the neighboring leg to remain in stance longer while the current leg completes its swing phase (an enforcement dictated by the first rule). 3. Reflexes for Adaptive Stepping Walking robots need the ability to perform extreme actions to adapt to severely irregular terrain and navigate obstacles. By initiating these actions via sensory cues and a cascade of reflexes [11], new, emergent, behaviors can be

5

implemented through the same neurobiologically-based walking control method, without the necessity of creating bolt-on, independent reflex actions. 3.1. Elevator Reflex The elevator reflex allows the robot to step over obstacles that are surmountable, but higher than the normal step height while walking [1], [2] (Fig. 2, left). Sensory information from increased loading of thorax-coxa (responsible for protracting and retracting the leg) and coxa-trochanter (responsible for levating and depressing the leg) joints can indicate the presence of such obstacles. Additional, hypothesized neural pathways in the model of the stick insect thoracic ganglia trigger a cascade of reflexes that raise the stepping height in order to bypass the hurdle. Due to limited capabilities of the leg microcontrollers, BILL-Ant-a uses a contact switch located on the front of each foot to detect collisions with obstacles. When a foot impacts an obstruction during its swing phase, joint directions are briefly altered to become retraction, levation, and extension actions. These actions retract the foot slightly from the face of the obstacle, in order to remove contact with the obstacle face, and employ a levating action with a greatly-increased muscle activation level. The higher levation activation propels the leg upward in an effort to clear the obstacle. After this brief alteration to the leg behavior the nominal, sensory-coupled joint actions resume. If the obstacle has been cleared, nominal joint actions continue; if not, the process repeats. 3.2. Searching Reflex When a leg steps into a hole or gap, the foot moves lower than the expected level of the ground [1], [2]; this triggers the searching reflex (Fig. 2, right) where the leg alters its movement pattern in order to locate a suitable purchase beyond the opening. As with the elevator reflex, this behavior is initiated by sensory cues (the lack of ground contact at the expected height), which causes a reflex cascade that elicits a searching rather than stepping behavior, that modifies the stepping actions within the existing framework of the walking controller. The searching reflex is initiated when the foot drops below a level where the ground is expected. As with the elevator reflex, the searching reflex results in the joint directions becoming retraction, levation, and extension. The slight retraction action, as before, clears the leg for executing a levation action with increased muscle activation. Since the foot is at a lower level than is normally

6

seen during stepping, the increased levation muscle activation promotes the returning of the foot to its nominal height during a step.

Figure 2. BILL-Ant-a performing an elevator reflex (left) and a searching reflex (right). For the elevator reflex (left) the front-right leg alters its nominal 0.5in (1.72cm) step height to clear the 0.79in (2.0cm) obstacle. The right image shows the middle-right leg in the middle of performing a searching reflex to clear the gap, while the middle-left leg is about to initiate the same reflex. The increased levation muscle activation is responsible for high stepping seen in the elevator and searching reflexes.

4. Goal Seeking and Modulation of Walking Control Straight walking is sufficient to test the implementation of stepping, gait generation, and reflex behaviors, but does not produce an interesting mobile robot. In order to make a walking robot useful, it needs the ability to turn. Also, a level of autonomy that allows the robot to internally decide its actions greatly improves upon its usefulness. A high-level head module is able to detect visual sensory information and, based on this information, actuates a neck motor holding the visual sensors and uniquely modulates the parameters of the leg controllers. This module does not micromanage or dictate the generation of stepping actions, but alters stepping parameters such that turning occurs, reminiscent of the way that the central body complex is thought to modulate the local controller in the thoracic ganglia in insects [12]. Unique modulations are used since each of the six legs requires different changes in stepping motion to perform a turning behavior. For example, front legs alter their path from the sagittal plane much more severely than do the hind legs. Also, legs on the inside of the turn diverge more from the sagittal plane than do outside legs. This alteration in stepping actions is accomplished by modulation of the joint direction transition values (leg load and joint angles responsible for influencing in which direction each leg joint should move) and the muscle activation values that determine joint speed. For example, bringing the anterior extreme position (AEP) and posterior extreme position (PEP) closer together

7

shortens the sweep of the thorax-coxa joint. Increasing the femur-tibia (FTi) joint angle setpoint that causes the FTi joint to change from flexion to extension, and similarly reducing the FTi joint angle responsible for causing a transition from extension to flexion promotes larger excursions of the FTi joint. Combining these new setpoints with decreased protraction and retraction muscle activations and increased extension and flexion muscle activations causes the leg to change from nominal stepping in the sagittal plane to angled stepping, which then induces turning in the robot. 5. Conclusion This paper demonstrates the successful implementation of a neurobiologicalbased control system in a physical hexapod robot. A distributed collection of leg control modules is responsible for the manipulation of individual leg joints, much like the function of thoracic ganglia. Leg load and joint angle data are used to coordinate joint movements such that stepping patterns are produced. And interaction with the environment causes dynamic alterations to those patterns, which allows the robot to navigate past obstacles through a pair of reflexes. Sharing a limited amount of information between these modules provides the ability for stable, speed-dependent gait generation to occur. A head module that processes data from a pair of neck-mounted light sensors provides the robot with the ability to perform phototaxis. This module uses visual information to both actuate the neck in order to locate a goal and then uniquely adjusts parameters on each of the six leg control modules, in a manner similar to the descending modulation function of the insect brain and subesophageal ganglion, which enable the robot to perform turning maneuvers. This collection of features and abilities represents the first implementation of such a neurobiological-based control system on a physical hexapod. References 1. P. Arena, L. Fortuna, M. Frasca, and L Patané, “Sensory Feedback in locomotion control,” in Dynamical Systems, Wave-Based Computation and Neuro-Inspired Robots, Vol. 500, Part II, pp. 143-158, 2008. 2. R.D. Beer, R.D. Quinn, H.J. Chiel, and R.E. Ritzmann, “Biologically Inspired Approaches to Robotics,” in Communications of the ACM, Vol. 40, No. 3, March 1997. 3. R. Brooks, “A Robot that Walks: Emergent Behaviors from a Carefully Evolved Network,” MIT AI Lab Memo 1091, February 1989.

8

4. H. Cruse, J. Dean, U. Müller, and J. Schmitz, “The Stick Insect as a Walking Robot,” in Proceedings of the Fifth Int. Conf. on Advanced Robotics. Vol. 2 , S. 936-940 – ICAR, (1991). 5. H. Cruse, T. Kindermann, M. Schumm, J. Dean, and J. Schmitz, “Walknet – a biologically inspired network to control six-legged walking,” in Neural Networks 11 (1998) 1435–1437. 6. Ö. Ekeberg, M. Blümel and A. Büschges, "Dynamic Simulation of Insect Walking", in Arthropod Structure & Development 33 287 (14 pages), 2004. 7. M. Frik, M. Guddat, M. Karatas, and C. D. Losch, “A novel approach to autonomous control of walking machines,” in Proc. of the 2nd Int. Conf. on Climbing and Walking Robots (CLAWAR1999), (ed. G. S. Virk, M. Randall & D. Howard), pp. 333– 342. Bury St. Edmunds, London: Professional Engineering Publishing Limited, (1999). 8. D. Hess and A. Büschges, “Sensorimotor Pathways Involved in Interjoint Reflex Action of an Insect Leg,” in Journal of Neurobiology 1997; 33(7):891-913. 9. W.A. Lewinger, B.L. Rutter, M. Blümel, A. Büschges, and R.D. Quinn, “Sensory Coupled Action Switching Modules (SCASM) generate robust, adaptive stepping in legged robots,” in Proc. of the 9th Int. Conf. on Climbing and Walking Robots (CLAWAR2006), pp. 661–671, Brussels, Belgium, Sept 12–14, 2006. 10. W.A. Lewinger, and R.D. Quinn, “BILL-LEGS: Low computation Emergent Gait System for Small Mobile Robots,” in Proc. of IEEE International Conference on Robotics and Automation (ICRA'08): pp. 251256, Pasadena (CA), USA, May 11-23, 2008. 11. L. Mu and R. E. Ritzmann, “Interaction between descending input and thoracic reflexes for joint coordination in cockroach. II Comparative studies on tethered turning and searching,” in The Journal of Comparative Physiology A, 194:299 – 312, (2008). 12. J. Wessnitzer and B. Webb, “Multimodal sensory integration in insects – towards insect brain control architectures,” in Bioinspiration and Biomimetics, Vol. 1, pp. 62 – 75, (2006).