Neuromorphic vision sensors for mobile robots Alan Stocker and Giacomo Indiveri Institute of Neuroinformatics, ETH/University Zurich, Zurich, Switzerland E-mail: [alan|giacomo]@ini.phys.ethz.ch Real-time vision-based vehicle navigation tasks are typically computationally in- tensive and significantly complex in terms of resources used. Moreover, mobile ap- plications of navigation systems place severe constraints on their size and power con- sumption. These constraints can be satisfied by using parallel image processing archi- tectures and parallel control algorithms implemented with analog VLSI circuits. In this paper we present two examples of neuromorphic VLSI sensors that process motion sig- nals to provide continuous time analog output signals that can be useful for real-time navigation tasks. Neuromorphic analog VLSI vision sensors are typically compact, low-power devices that implement parallel architectures to perform in real-time, im- age (pre)processing operations and provide results that would otherwise require CPU intensive computations [5]. The two neuromorphic sensors we propose in this paper measure respectively the scene’s 2D optical flow field and the target’s position and its direction of motion in one dimension. 1 2-D Optical flow sensor The types of operations performed by this sensor, the computational elements used and their architectures are very similar to the ones found in biological visual systems. Na- ture demonstrated remarkable examples of efficient motion sensing systems: The per- formance of e.g. the fly’s visual system in terms of weight, size, power consumption and speed cannot be achieved by any artificial system yet. The main advantages of the biological systems are based on the parallel network structure, band-pass filtering and adaptation on several processing levels which is in high contrast to traditional com- puter vision processing. We present a focal-plane aVLSI chip that performs smooth optical flow computation in two image dimensions. The chips contains a array of interconnected motion units and has been reported in its preliminary form earlier [9]. The architecture of the system is such that the derived local velocity output is the minimal error solution of the intersection of constraints within a weighted neighbor- hood. This is achieved by having spatially distributed competitive computation in two recurrently connected networks. The tunable size of the neighborhood-kernel reflects the strength of the a priori assumption that motion cells with a receptive field close to each other are more likely to see the same object. As such one can choose also the 1