Robotics Research Lab
USC Computer Science
USC Engineering
/ Research / Projects / Adaptive Spatio-Temporal Organization in Multi-Robot Systems

Objectives Problem Domain Controller Architecture Cooperative Strategies
Experiments Results Publications Support

Research Objectives

This project looks at developing general adaptive capabilities in robots and multi-robot systems. By general capabilities we mean capabilities that perform well over a range of tasks and environments. Such capabilities go further than specialized skills in increasing the autonomy and applicability of robotic systems.

One form of general adaptivity is spatio-temporal adaptivity. This project uses reinforcement learning to allowing robots to dynamically adjust their behavior to any given environment while performing a set task.

Transportation - An Initial Problem Domain

To best demonstrate how the structure of a group of robots can adapt, we look at transportation, where a group of robots have to traverse an environment between a source and a sink. Transportation is a sub-problem of foraging in that the positions of the sources and sinks are known, eliminating issues of search. The simplicity of transportation makes it easier to show that the improvements in the group performance are due to a change in the groups spatio-temporal structure.

Controller Architecture

The controllers we use for the robots are behavior-based. The robots used in the experiment below had three different abstract, high-level behaviors: direct target approach, wall following, and stopping. By learning how to relate different input states to any of these three behaviors, the robots can improve their performance. On top of the movement behaviors we implemented a concurrent set of communication behaviors for cooperation. The communication strategies are described in a separate section below.


The three high-level movement behaviors take care of lower level tasks such as obstacle avoidance, position estimation, and keeping track of what target to visit next, the source or the sink.

The Adaptive Control Algorithm

The adaptive element of the controller uses reinforcement learning to relate the different input states to the available movement behaviors. The reinforcement learning uses a naive Q(λ) algorithm with an accumulative eligibility trace and a Boltzmann Softmax action selection algorithm.

The input space is made up of several highly abstracted virtual sensors. In the experiment we describe below we used a five bit input vector. The first bit indicated the presence of a visible target. The second bit was on for a relatively short amount of time at random intervals, providing an unbiased fragmentation of the input state. The three final bits reflected the presence of the three colors used for the color markings. They only reported on the markings on the closest of the visible robots.

Cooperative Strategies

To improve the learning rate of the group we developed two communication strategies based on language-less communication in animals and humans. The first strategy was called expressed behavior evaluation and the second imitation. In order to scale to large groups, the strategies only used local communication.

The expressed behavior evaluation strategy was inspired by wordless calls of encouragement or warning in humans and animals. Humans and some animals can recognize the active behavior in others and estimate their input state. This is very difficult to achieve in robots, so our robots broadcast locally every new state-action pair they went through. When a robot received a state-action message referring to a Q-table entry with an exceptional value in the robot's own Q-table, the robot would send a feedback message to the originator of the state-action message. On reception of a feedback message, a robot would assume that its current state and action was an exceptional combination and update its Q-table accordingly.

The imitation strategy was inspired by imitation learning where a student repeats the actions of a teacher, hence experiencing the same state-action chains and receiving similar reward. In our system this is implemented as storing and broadcasting locally the eligibility trace that lead to the highest reward. Similar cooperation strategies have been shown to work in a grid-world, pursuit domain by Ming Tan. Our results show that the same improvements in learning rate can be achieved in more realistic simulations.


As an initial demonstration of spatio-temporal adaptation, we did a series of experiments where five Pioneer 2DX robots simulated using Player and Stage , had to traverse a 6-by-7 meter rectangle containing a source and a sink on opposite short sides. The robots were equipped with PTZ cameras and SICK laser range finders. The robots also wore color markings emphasizing their orientation with red on the left side, green on the right, and yellow at the rear.


The robots received a reward on arrival at a target, source or sink. The reward, r, was inversely proportional to the traversal time, t, and relative to the optimal traversal time, to: r=1+(to/t).


Our experiments demonstrated statistically significant improvements in both individual and group performance. They also demonstrated the value of our cooperative strategies. Below a graph illustrates the performances of the three different learning strategies.


The improvements were due to both spatial and temporal reorganization. The robots learned to follow a circular path, minimizing interference. They also learned to stop when they saw other robots approaching, i.e. they did not see their yellow markings.

Support The project is funded in part by the Deparment of Energy (DOE), Robotics and Intellgent Machines (RIM), grant number DE-FG03-01ER45905, for "Multi-Robot Learning in Tightly-Coupled, Inherently Cooperative Tasks", and in part by the Office of Naval Research (ONR), Defence University Reseach Instrumentation Program (DURIP), grant number 00014-00-1-0638 for "Equipment Support for Dynamic Adaptive Wireless Networks with Autonomous Robot Nodes".