A Framework for Controlling Multi-Robot Systems Using Bayesian Optimization and Linear Combination of Vectors

Stephen D. Jacobs, R. Michael Butts, 裕二 池谷, Ali Baheri, Guilherme A. S. PereiraarXiv 2022 · DOI · arXiv

This paper, authored by researchers at West Virginia University, presents a unified approach to a classic problem in robotics: how do we get a group of independent robots to work together without a "central brain" telling them exactly what to do?

TL;DR

The researchers propose a generalized, optimizable framework that converts raw sensor data into robot velocities by treating control as a linear combination of vectors weighted by scalar measurements. By using Bayesian Optimization to tune these weights, they successfully demonstrated that a single mathematical structure can handle diverse tasks like flocking, pattern formation, and foraging.

What Problem Does This Paper Address?

In decentralized multi-robot systems, each robot must make decisions based only on what it can see or hear locally. Traditionally, researchers have designed "bespoke" controllers for specific tasks—one for collision avoidance, another for flocking, and a third for searching.

The problem is that these specialized controllers don't play well together. If you want a robot to avoid a wall while staying with its flock, combining those two distinct algorithms is mathematically messy and often requires tedious manual parameter tuning. The authors sought a "universal language" for these controllers that is both flexible enough for many tasks and easy to optimize automatically.

Methods

The core of the paper is a deceptively simple mathematical framework. They classify all robot sensor data into two categories:

  1. Scalars ($S$): Numerical values like the distance to the nearest neighbor or the number of agents in a group.
  2. Vectors ($\vec{V}$): Directional data like the unit vector toward a goal or the heading of a neighbor.

The control velocity ($\vec{V}_c$) for a robot is calculated using a matrix of parameters ($P$) that multiplies the scalars to create weights for the vectors: $$\vec{V}_c = (P \times S) \cdot \vec{V}$$

Key Innovation: Bayesian Optimization (BO)

Because the relationship between these parameters and the final "success" of a swarm is complex and "black-box," the authors use Bayesian Optimization. Instead of a human guessing the values in matrix $P$, BO treats the simulation as an expensive function to evaluate. It builds a probabilistic model (a Gaussian Process) of the parameter space and intelligently picks the next set of parameters to test, balancing "exploration" (trying new things) and "exploitation" (refining known good areas).

Datasets

As is common in control theory research, the "datasets" consist of simulated environments and experimental setups used to validate the framework:

  • Simulated Environments: Custom-built scenarios for flocking (40 agents), cohesion/segregation (3 groups of 5), radial pattern formation (100 agents), and a 36-point grid for searching/foraging.
  • Real-World Validation: A laboratory setup using two holonomic (multi-directional) robots tracked by a Vicon motion capture system to perform an "antipodal swap" (robots switching places without colliding).

Key Insights and Ideas

  • Linearization is Sufficient: While the "perfect" control law for a swarm might be highly non-linear, the authors show that a linear combination of the right features (scalars and vectors) is expressive enough to generate complex, emergent behaviors.
  • Feature Engineering Matters: The authors found that using the inverse of a distance (e.g., $1/d^3$) was much more effective for repulsion behaviors than using the raw distance. This "intuition" helps the linear framework mimic non-linear physical forces.
  • Automated vs. Manual Tuning: For complex tasks like collision avoidance, Bayesian Optimization found better parameters than a 10,000-iteration Monte-Carlo (random) search, proving that the framework is highly "tunable" without human intervention.

Key Results

  • Collision Avoidance: In an antipodal swap test, the BO-optimized controller achieved a cost of 253.4, which was lower than any of the 10,000 random trials (the best of which was significantly higher).
  • Scalability: The framework successfully managed a radial pattern formation with 100 agents, showing that the decentralized nature allows it to scale to large swarms without a computational bottleneck.
  • Versatility: The same underlying equation was used to produce five distinct behaviors, ranging from biological-style flocking to industrial-style grid searching.

Limitations and Future Work

The authors identify two main hurdles for the future:

  1. Reward Shaping: To use Bayesian Optimization, you need a "cost function" (a mathematical way to say if the robots did a good job). Defining this function is still difficult and can lead to unintended robot behaviors if not designed carefully.
  2. The Curse of Dimensionality: If you give the robot too many scalars and vectors to consider, the search space for the optimization becomes too large. Future work will look into "structure selection"—automatically determining which sensors are actually important for a specific task.
  3. Theoretical Proofs: While the system works in practice, the authors aim to provide formal proofs that Bayesian Optimization will consistently find the global optimum for these types of multi-robot control laws.

Summary updated

Abstract

We propose a general framework for creating parameterized control schemes for decentralized multi-robot systems. A variety of tasks can be seen in the decentralized multi-robot literature, each with many possible control schemes. For several of them, the agents choose control velocities using algorithms that extract information from the environment and combine that information in meaningful ways. From this basic formation, a framework is proposed that classifies each robots' measurement information as sets of relevant scalars and vectors and creates a linear combination of the measured vector sets. Along with an optimizable parameter set, the scalar measurements are used to generate the coefficients for the linear combination. With this framework and Bayesian optimization, we can create effective control systems for several multi-robot tasks, including cohesion and segregation, pattern formation, and searching/foraging.
BibTeX
@misc{jacobs2022framework,
  author = {Stephen D. Jacobs and R. Michael Butts and 裕二 池谷 and Ali Baheri and Guilherme A. S. Pereira},
  title = {{A Framework for Controlling Multi-Robot Systems Using Bayesian Optimization and Linear Combination of Vectors}},
  year = {2022},
  eprint = {2203.12416},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2203.12416},
}
Figure