Evolution of spiking neural circuits in autonomous mobile robots

We describe evolution of spiking neural architectures to control navigation of autonomous mobile robots. Experimental results with simple fitness functions indicate that evolution can rapidly generate spiking circuits capable of navigating in textured environments with simple genetic representations that encode only the presence or absence of synaptic connections. Building on those results, we then describe a low-level implementation of evolutionary spiking circuits in tiny micro-controllers that capitalizes on compact genetic encoding and digital aspects of spiking neurons. The implementation is validated on a sugar-cube robot capable of developing functional spiking circuits for collision-free navigation. I. Spiking Neural Circuits The great majority of biological neurons communicate using self-propagating electrical pulses called spikes. Computational approaches to the study of brain function define two classes of neuron models that, among other things, differ in their interpretation of the role of spikes. Connectionist models [23], by far the most widespread, assume that what matters in the communication is the firing rate of a neuron, that is, the average quantity of spikes emitted by the neuron within a relatively long time window (for example, over 100 ms). In these models the real-value output of a neuron represents the firing rate, possibly normalized relatively to the maximum attainable value. Pulsed models [19], instead, are based on assumption that the firing time, that is, the precise time of emission of a single spike, may convey important information [25]. Often, these pulsed network models use complex activation functions that represent the emission of spikes on a very fine timescale [22]. Leaving aside the question of whether information transmitted among neurons is encoded by firing rate, firing time, or a combination of both, artificial spiking neural networks are attracting increased attention because they could capture and exploit more efficiently (i.e., with fewer neurons or with higher probability) non-linear time series of input signals; can be implemented in tiny and low-power chips [13] that exploit the sub-threshold physics of transistors in analog VLSI [20]; and allow biologically plausible investigations of computation in nervous systems. In this paper we are concerned mainly with the latter issue and show that adaptive networks of spiking neurons can be efficiently implemented also in tiny, low-cost, and largely available digital circuits. February 17, 2005 DRAFT INTERNATIONAL JOURNAL OF INTELLIGENT SYSTEMS, VOL. XX, 2005 102 Designing circuits of spiking neurons that display a desired functionality is still a challenging task. The most successful results in the field of robotics obtained so far focused on the first stages of sensory processing and on relatively simple motor control. For example, Indiveri et al. [12] developed neuromorphic vision circuits that emulate interconnections among neurons in the early layers of the biological retina in order to extract motion information and implement a simple form of attentive selection. These vision circuits have been interfaced with a Koala robot and their output has been used to drive the wheels of the robot in order to follow lines [14]. In another line of work, Lewis et al. developed an analog VLSI circuit with four spiking neurons capable of controlling a robotic leg and adapting the motor commands using sensory feedback [18]. This neuromorphic circuit consumes less than 1 microwatt and takes less than 0.4 square millimeters of chip area. Despite these promising implementations, there are not yet methods for developing complex spiking circuits that could display minimally-cognitive functions or learn behavioral abilities through autonomous interaction with a physical environment. Artificial evolution thus may represent a promising methodology to generate networks of spiking circuits with desired functionalities expressed as behavioral criteria (fitness function). In previous work [4], we showed that evolution of spiking circuits can generate functional networks of spiking circuits for vision-based navigation of autonomous robots. Neuro-ethological analysis of an evolved circuit revealed functional specialization of single neurons and the role of spiking correlation on behavior. More recently, DiPaolo [3] used a similar approach to investigate the role of noise and synaptic plasticity in light-directed tasks. In this article, we expand our previous work [7], [4] and describe a compact digital implementation of evolutionary spiking circuits that capitalize on our findings that such circuits do not require specification of synaptic weights and thus result in compact genetic encodings. The resulting evolutionary spiking circuit on chip, which occupies less than 50 bytes of memory, is validated on a sugar-cube robot that autonomously and reliably develops the ability to navigate around a maze in a less than an hour. A preliminary implementation of this model was described in [7]. Here we describe a final implementation, a new set of experiments, and the analysis of evolved network architectures. In the next section we describe the network architecture and genetic representation used February 17, 2005 DRAFT INTERNATIONAL JOURNAL OF INTELLIGENT SYSTEMS, VOL. XX, 2005 103 in these experiments. In the section that follows we briefly describe a set of evolutionary experiments on vision-based navigation with a neuron model that captures non-linear dynamics of synaptic integration and post-spike membrane behavior. These experiments are based on the specifications that we presented in [4]. We then describe the implementation in a micro-controller of a simplified neuron model and evolutionary algorithm and present a set of evolutionary experiments with a fully autonomous sugar-cube robot. Finally, we discuss the relationship between our low-level digital implementation and other analog VLSI implementations of spiking networks, as well as scalability issues and extensions of