The control of locomotion takes place at three levels: posture (stabilization and orientation), steering (obstacle avoidance and goal-seeking), and navigation (taking routes to unseen goals). Here I explore three strategies for steering. The Pictorial Strategies use splay and image height information and are well-known. The Heading Strategy, as proposed 40 years ago by J. J. Gibson, places the perceived heading on gaps and goals in the environment. A third strategy is the Equalization Strategy---steering to equate the magnitude of optic flow on each side of the heading. It has recently been found that bees use this latter strategy and it has also been implemented in robots. I present experiments testing for the presence of each of these strategies in a long corridor, a corridor with an obstacle and a target, and situations with only a target. The patterns of path data indicate that all three strategies are engaged by subjects. I then model those conditions where the Equalization and Heading Strategies appear to dominate. A simulated agent that can detect optic flow is run through the same conditions as the subjects. If the agent simply sums the outputs of each strategy, it follows paths similar to the subjects in a number of experiments. These are some of the first experiments indicating that humans do actually use optic flow to steer.
This is the final word, so to speak, on all the robot stuff I've done. It summarizes the work from the two conference papers below and discusses more of the simulations that I've done and the transfer of the avoidance and tag behaviors to Ramona.
Abstract There are striking parallels between ecological psychology
and the new trends in robotics and computer vision, particularly regarding
how agents interact with the environment. We present some ideas from ecological
psychology, including control laws using optic flow, affordances and action
modes, and describe our implementation of these concepts in a small mobile
robot which can avoid obstacles and chase or flee moving targets solely
using optic flow. The properties of these methods were further explored
in simulation. This work ties in with that of others arguing for a methodological
approach in robotics which foregoes a central model/planner. Ecological
psychology may not only contribute to robotics, but robotic implementations
may in turn provide a test bed for ecological principles and a source of
ideas which could be tested in animals and humans.
Be warned that the margins on this are a little too wide so you can't see all of it in ghostview. It does print out okay though, on some printers (not DEC's)
ABSTRACT: There are striking parallels between ecological psychology and new trends in robotics and computer vision, particularly regarding how agents interact with the environment. We present some ideas from ecological psychology, including control laws using optical flow, affordances and actions modes, and describe our implementation of these concepts in a small mobile robot which can avoid obstacles and play tag solely using optical flow. This work ties in with those of others arguing for a methodological approach in robotics which foregoes a central model/planner. Ecological psychology may not only contribute to robotics, but robotic implementations in turn provide a test bed for ecological principles and sources of ideas which could be tested in animals and humans.
Investigative Opthalmology and Visual Science Abstract Book, 37(3), S515.
ABSTRACT: Purpose. We compare the contribution of various sources of information to the control of steering, including optical expansion, disparity, and elevation angle. Methods. Subjects navigated through a slalom course made up of pairs of vertical posts, or "gates." A joystick controlled their x-position while z-velocity was held constant to equate duration. Gates were placed at the extrema of sum of sines curves, and about 5 were visible at one time. Display factors were: (1) Expansion: expanding textured gates vs. untextured gates of constant width, (2) Stereo: stereoscopic vs. monoptic viewing, (3) Ground: gates on an implicit ground plane vs. extending to the bottom of the screen. Dependent measures included the percentage of gate collisions and the proportion of total power at the course frequencies. Results. For percentage of collisions, there were main effects of Expansion (18% of total variance) and Stereo (6%), but not Ground. Once one source of information was available, no further improvement occurred until the full-cue condition. Power spectra revealed significant effects of Expansion (28%), Stereo (10%) and Ground (11%), but there was no improvement with combined information, even in the full-cue condition. Conclusions. These three sources of information are each sufficient and substitutable. There appear to be no additive effects, except in the full-cue case for collision avoidance. Thus, steering can be controlled in the expansion condition without traditional depth cues; ordinal depth information may indicate which gate to steer toward next. Supported by NIH EY10923 and an NSF Graduate Fellowship.
ABSTRACT: Some recent work with autonomous robots has focused on using optical flow for "direct" control of speed and rotation in obstacle avoidance and other simple behaviors. This work has been inspired by work with insects showing similar mechanisms. To extend these behaviors, three methods of maze navigation are investigated in a simulated robot modeled after a real one. A motor-based method places biases in the obstacle avoidance control law used previously. A perception-based method uses optical flow to detect possibilities for action (e.g., to turn left or right). Both of these require that the agent have a list of biases in order to navigate. The third method, called the Salience Centroid Method, is based on a theory of the role of the hippocampus in rat navigation. This method trades off the memory of the first two for more advanced perceptual processing and allows the most flexible behavior.
Investigative Opthalmology and Visual Science Abstract Book, 38(4), S79.
ABSTRACT: Purpose. When locomoting through a complex environment, do people use on-line control or plan a path ahead? Subjects steered through a slalom course of "gates" with a joystick while the visible band of gates and its preview distance were manipulated. Methods. Gates (pairs of vertical posts) were randomly positioned about a straight path on a textured ground plane. Z-velocity was constant (1.25 or 2.5 eyeheights/sec) while the joystick controlled X-velocity, and hence heading. Only a few gates were visible ahead and then disappeared before the observer passed them. We varied the visible band or number of visible gates (1, 2 or 3) and the preview distance or number of invisible gates between the visible band and the observer (0, 1 or 2). Successful passage through a gate was indicated by a "beep". Dependent variables included constant and variable errors in the passage point and in the position of the turn with respect to each gate. Results. Performance was high with 1 gate in the visible band, whereas 2 or 3 visible gates added little improvement. Performance was best witha preview distance of 0 gates, declined with 1 invisible gate and deteriorated markedly with 2 invisible gates. There were no effects of speed, indicating that the results are due to the number of gates rather than time. Conclusions. Subjects generally control steering on-line with respect to the next visible object. A view of additional objects does not provide much advantage, consistent with a lak of planning ahead. Occluding objects before passage reduces performance; subjects can delay adjustments for 1 invisible object with reasonable accuracy, but cannot "pipeline" information or plan for 2 invisible objects. Supported by NIH EY10923. None.