Group Behaviors

[Group Behaviors Image]


Birds, fish, and many other animals are able to move gracefully and efficiently as a herd, flock, or school. We would like to reproduce this behavior for herds of artificial creatures with significant dynamics. To run as a herd, animals must remain in close proximity while changing direction and velocity and while avoiding collisions with other members of the herd and obstacles in the environment. We have explored the performance of a control algorithm for modulating the motion of each individual in a herd of dynamically simulated legged robots.

In contrast to most previous implementations of algorithms for group behaviors, we are using this algorithm to control a robot herd where the individuals have significant dynamics. The problem of controlling the robot herd more closely resembles that faced by biological systems because of the underlying dynamics of the individuals in the herd. Each robot in the herd is a dynamic simulation of a physical robot and a control system. As such, the robots have limited acceleration, velocity, and turning radius. Furthermore, the control algorithms are inexact, resulting in both transient and steady-state errors in velocity control. Required changes in velocity are delayed by as much as half a running step because the control system can influence velocity during only the flight phase of the running cycle. To understand the effect of the underlying dynamics, we compared the performance of the herding algorithms on the robots with full dynamics and on particle systems with perfect velocity control.

The herding algorithm computes a desired velocity for each individual based on the location and velocity of its visible neighbors. This desired velocity is then used as an input to the locomotion control system for the robot. We compared the performance of this algorithm on a herd of point-mass objects and a herd of dynamically simulated running robots for a test suite of four problems: steady-state motion, acceleration and deceleration, turning, and avoiding obstacles. For this test suite, all individuals in the herd of robots remained upright and only a small number of collisions occurred. The performance of this herd was not as robust as that of the point-mass system because of the dynamics of the individual robots.

Algorithms for high-level behaviors of dynamic simulations are needed for the construction of synthetic actors with robust and realistic motion that can respond interactively to changes in the environment. A dynamic simulation in concert with a control system will provide natural looking motion for low-level behaviors such as walking, running, and climbing. High-level behaviors such as obstacle avoidance, grouping, and rough terrain locomotion will allow the actor to function in and interact with a complex and unpredictable environment.

Publications

Project Members