Dynamic Heterogeneous Robot Teams Engaged in Adversarial Tasks

M. Bernardine Dias, Brett Browning, Manuela Veloso, Anthony StentzAsian 2018 · DOI

Abstract

As we progress towards a world where robots play an integral role, a critical problem remains: the formation of dynamic, heterogeneous robot teams where little information is known a-priori about the tasks, the robots, or the environments. This paper defines the "Pickup Team Challenge"—the goal of enabling diverse, unfamiliar robots to form a cohesive team and execute complex, adversarial tasks in under 10 minutes. We propose a framework that merges market-based task allocation (TraderBots) with play-based synchronization (STP) to allow robots to "plug and play" into a team without extensive prior hand-engineering. This system allocates roles dynamically and executes synchronized actions to accomplish complex objectives in uncertain environments.

Introduction: The Pickup Team Challenge

In traditional multi-robot research, teams are typically "pre-season" squads. Engineers spend months hand-tuning coordination for a specific set of robots to perform a specific task in a controlled environment. While effective in the lab, this approach fails in real-world scenarios for three primary reasons:

  1. Scalability: It is prohibitively expensive for a single organization to maintain every type of robot required for every possible task.
  2. Emergency Response: In disaster scenarios, there is no time to code custom coordination logic; responders must use whatever robotic assets are immediately available on-site.
  3. Robustness: If a robot fails or a new asset is introduced mid-mission, traditional hard-coded coordination often collapses.

The "Pickup Team Challenge" envisions a world where robots from different manufacturers, running different internal algorithms, can meet for the first time and effectively collaborate on a task—such as a "treasure hunt" against an adversary—with minimal setup time.

Methodology: Merging Markets and Plays

The proposed architecture combines two established technologies to bridge the gap between high-level coordination and low-level execution:

TraderBots (Task Allocation)

TraderBots is a market-based coordination mechanism where robots are treated as self-interested agents in a virtual economy. Robots "auction" tasks to one another, bidding based on their internal costs (e.g., battery life, distance, or specific sensor capabilities). This decentralized approach ensures robust task allocation without a single point of failure.

STP: Skills, Tactics, and Plays (Execution)

Originally developed for RoboCup soccer, the STP framework provides a hierarchy for team execution:

  • Plays: Fixed team-wide plans (e.g., a specific search pattern).
  • Tactics: Individual roles within a play (e.g., "The Scout" or "The Mapper").
  • Skills: Low-level control algorithms (e.g., "Move to coordinates X, Y").

The Innovation: Market-Based Role Assignment

The core innovation of this work is using TraderBots to negotiate roles within an STP Play. Instead of hard-coding which robot performs which tactic, the robots use a market auction to decide who is best suited for a role based on their real-time state and capabilities. This decouples the tactic from its implementation; a robot does not need to know how its teammate performs a task, only that the teammate has the capability to do so.

Experimental Setup: The Treasure Hunt

The framework was evaluated using a physical "Treasure Hunt" scenario involving a diverse fleet of hardware:

  • Pioneer Robots: Equipped with SICK laser range finders for mapping and localization.
  • Segway RMP Robots: High-speed platforms with cameras for visual object identification.
  • Sony AIBO Robots: Small legged robots capable of navigating tight spaces.
  • Gator Robots: Large autonomous vehicles intended for outdoor expansion.

To handle adversarial elements, the system employs a Gibbs distribution to select plays. This introduces a level of stochasticity—randomizing play selection while weighting successful strategies more heavily—which prevents opponents from easily predicting and countering the team's behavior.

Key Insights and Results

The system demonstrated several critical capabilities for dynamic team formation:

  • The 10-Minute Rule: The authors established a benchmark that teams should organize and begin a task within 10 minutes of initial contact.
  • Heterogeneity as a Feature: The framework successfully paired robots with complementary strengths. For instance, a Pioneer robot (strong localization) could provide a map to a Segway robot (strong vision) to facilitate treasure discovery.
  • Dynamic Role Flexibility: By utilizing the TraderBots market, the team could re-allocate roles on the fly if a robot’s battery ran low or a sensor failed, requiring no human intervention.
  • Play Adaptation: The team demonstrated the ability to adjust the "weight" of specific plays based on success or failure, effectively learning which strategies were most effective against specific obstacles.

Limitations and Future Work

While the framework provides a solid foundation, several challenges remain. There is a pressing need for a universal Task Specification Language to allow robots to describe their capabilities to "strangers" in a standardized format. Additionally, while the robots can negotiate, they do not yet "learn" the specific reliability or quirks of their teammates over time. Future research will focus on incorporating teammate modeling and transitioning these capabilities from controlled indoor environments to complex, unconstrained outdoor settings.

Summary updated

Abstract

As we progress towards a world where robots play an integral role, a critical problem that remains to be solved is that of dynamically formed heterogeneous robot teams where little information is known a-priori about the tasks, the robots, and the environments in which they will operate. We define this challenge as the Pickup Team Challenge. Successful solutions to forming pickup teams will enable researchers to experiment with larger numbers of robots; beyond what they can support and maintain. Additionally, enabling such teams will have a large impact on the ability of industry to efficiently and cost-effectively integrate new robot technology with existing legacy teams. In this paper, we define the challenge of pickup teams and relate its importance to multi-robot research. In our prior work, we have developed techniques for effective collaboration using marketbased techniques and for synchronizing team activity through plays. We build on these prior approaches to move towards a complete system that is able to allocate roles amongst robots in a pickup team, and to execute synchronized team actions to accomplish a complex task.
BibTeX
@inproceedings{dias2018dynamic,
  author = {M. Bernardine Dias and Brett Browning and Manuela Veloso and Anthony Stentz},
  title = {{Dynamic Heterogeneous Robot Teams Engaged in Adversarial Tasks}},
  booktitle = {Asian},
  year = {2018},
  doi = {10.1184/r1/6554621},
}
Figure