A neural model of visually guided steering, obstacle avoidance, and route selection

A neural network model is developed to explain how humans can approach a goal object on foot while steering in such a manner as to avoid collisions with obstacles in a cluttered environment. Successful steering requires a dynamical interaction between a person's perceived heading direction, or direction of self-motion, and the egocentric locations of the goal and obstacles. Recent psychophysical research suggests that a goal acts as an attractor of heading, while obstacles act as repellers of heading. The Steering, Tracking, And Route Selection (STARS) model presented here combines neural representations of heading with goal and obstacle positions to produce realistic steering behavior. The STARS model extracts the heading direction from an optic flow field using network layers that model properties of cells in cortical areas MT and MST, and it constructs goal and obstacle representations by detecting and grouping motion discontinuities within the optic flow field. The STARS model also contains a circuit that maintains fixation on the goal object during locomotion by generating a smooth pursuit eye rotation. Rotating the eye during locomotion introduces systematic distortion of the optic flow field, and the STARS model corrects for the effects of eye rotation using extra-retinal signals. The STARS model's architecture captures the attractor/repeller dynamics of steering behavior, while clarifying the role of heading perception and eye movements in complex steering tasks. Computer simulations are used to demonstrate the model's performance on a number of steering tasks, including approaching goals at different distances and initial viewing angles, steering in the presence of single obstacles, and route selection tasks involving paths that cross in front of or behind obstacles. Simulation results are quantitatively compared with experimentally observed steering trajectories. The computer simulation system exploits the programmable capabilities and massive parallelism of modern 3D computer graphics processing units (GPUs) to achieve near real-time simulation performance. The STARS model is implemented in its entirety to execute on commodity graphics hardware within the general-purpose GPU simulation system.