Project Objectives
Develop complete, effective and scalable software for autonomous robot teams. Demonstrate robot teams with integrated perception, reasoning, learning, communication and cooperative strategies that solve complex multiagent tasks.
Project Website
Approach
We focus on the key challenges for building individually skilled autonomous robots and successful multirobot teams. These challenges include: localization, cooperative object tracking using communication and fusion of distributed sensing, single- and multi-agent learning, graceful degradation in the face of reduced knowledge, team coordination, and demonstration on relevant heterogeneous platforms.
COAL: Cooperative Observation and Localization
It is clear that multiple cooperating robots can leverage and
multiply their sensing abilities to track moving objects and to localize
themselves in a dynamic environment. We believe it is also
clear that these problems (tracking and localization) are best
solved together in a unified framework because solutions to each part of
the problem improve solutions to the other parts.
We have been developing algorithms for COAL and evaluating and demonstrating
them on our robots.
Single- and Multi-agent Learning
Most proven learning methods involve a single agent learning to
act in a "Markovian" environment. In practice however, real-world
agents must learn and act with multiple other agents (either adversaries
and/or teammates) in a world that violates the Markovian restriction.
We have been investigating and evaluating techniques to enable our
agents to learn under these conditions.
Graceful degradation in the face of reduced knowledge
Robots acting in the real world are often faced with varying degrees
of information about their environment. For example,
sometimes they know very precisely
their location, other times they are lost. It is important that
the robots act appropriately in all these situations. Our work
in this area has focused on developing behaviors, called
``multi-fidelity behaviors'' that enable the agent to act appropriately
regardless of the confidence in their confidence in sensing.
Team Coordination
In addition to reliable individual behavior, inter-agent coordination
is crucial to overall team performance. We are investigating methods,
using communication and appropriate social control laws
to directly improve the performance of our
autonomous robot teams.
Demonstration on Relevant Heterogeneous Platforms
We have developed two new small (18cm) robotic platforms for
demonstrations of our research. Additionally, we have substantially
improved our existing Minnow (50cm) platform and acquired new
legged robots from Sony.
Recent Accomplishments
In the area of cooperative object tracking we have developed a new algorithm for communicating and fusing multiple simultaneous observations of objects for multiple robot teams. This work has been validated on teams of up to 4 robots observing a moving target.
We have developed and evaluated a new reinforcement learning technique called "Sequence Compression" that enables standard reinforcement learning algorithms like Q-learning to work effectively in non-Markovian domains.
We have implemented and demonstrated team coordination using communication on a team of autonomous soccer robots. The robots coordinate in several ways: 1) by sharing information about the location of adversaries and the soccer ball; 2) by sharing information about their own location to reduce interference; 3) by communicating about their intentions. Communication about intention (e.g. "I have the ball") enables the robot team to dynamically adjust to new situations. One example involves coordination between a goalie robot and a half-back. When the goalie leaves the goal area, it communicates this to the half-back so that it may fill in the missing role.
Plan
Distributed Fusion of Information
Teams of communicating robots offer a fascinating opportunity to
investigate the question of fusion of information. Building on our
current developments, we will research on new multi-robot localization
algorithms that are capable of using as input perception information
from different sources. This multi-robot localization will build a
probabilistic world model that will represent the beliefs of the
positions of the multiple robots with an accuracy that will
potentially increase with the accuracy of the distributed information
fused.
Learning to Create Efficient Task-Dependent Heterogeneous Robot Teams
Within our MARS research, we have developed and used a variety of
different robotic platforms. We have a current total of more than 20
fully functional robots. One interesting and challenging research
aspect that we face is the fact that our operational robots are not
all the same. Indeed our robots are of different types and have
different capabilities in autonomy, perception, and motion. In terms
of autonomy, some robots are computer-remotely controlled and others
are fully autonomous. In terms of perception, some sense the world
through vision only, some have additional sensors, and some others can
communicate to other robots. In terms of motion, we have robots with
engineered mechanical design that allows for highly accurate
dead-reckoning, other robots have differential drive, others are
omni-directional, and others are four legged robots. In FY02, we will
actively research on the question of how to combine our set of
heterogeneous robots to accomplish tasks that require distributed
capabilities among the robots. We will follow an approach in which our
robots will learn to improve their performance as an heterogeneous team
from experience and feedback from the environment.
Learning to Combine Planning and Reaction in Teams of Robots
We continue to pursue the investigation of the question of how to
effectively combine lookahead planning with reactive response in
dynamic and uncertain environments. This question is of particular
relevance in teams of robots as individuals robots need to make
decisions about the environment but they also need to reason about the
other robots in the team. We will introduce different levels of
dynamics in the environment and team, and we will investigate
algorithms to learn to control the tradeoffs between deliberation and
reaction.
Technology Transition
All software will continue to be available on the net. Our TeamBots software is already available online and is used by a number of researchers world-wide. The following robot designs and software funded under the MARS program have been developed by us and released to the community:
Relevant Publications