Adaptive Locomotion Control in Modular Robotics

Alexander Sproewitz, Rico Moeckel, Jérôme Maye, Masoud Asadpour, Auke Jan IjspeertInfoscience 2007

TL;DR

This paper introduces YaMoR, a modular robot platform, and a novel approach to adaptive locomotion control using Central Pattern Generators (CPGs) combined with a gradient-free optimization algorithm. This method allows modular robots to learn efficient and robust gaits in real-time, even with imperfect communication.

What Problem Does This Paper Address?

Generating robust and adaptive locomotion for articulated robots, especially modular ones, is a significant challenge. Modular robots, built from identical units, often result in sub-optimal actuator placement and under-actuation, requiring finely tuned locomotion patterns to compensate. Furthermore, the communication between a high number of modules in complex configurations needs to be reliable, even in the face of potential mechanical or electronic failures. Existing methods often struggle with the dynamic nature of modular robots and the complexities of real-world communication imperfections like delays and packet loss. This paper aims to address these challenges by developing a system that can learn and adapt locomotion patterns in real-time for modular robots, overcoming communication limitations.

Methods

The core of the proposed solution lies in a combination of two key techniques:

  • Central Pattern Generators (CPGs): Inspired by biological systems, CPGs are neural networks that produce coordinated rhythmic activity for locomotion. In this work, a CPG is implemented as a system of coupled amplitude-controlled phase oscillators, with each YaMoR module running one oscillator. These CPGs are designed to be flexible, allowing for adjustments in speed, direction, and gait type by altering parameters like amplitude, offset, and phase bias between oscillators. The CPG's inherent stable limit cycle and synchronization properties make it robust to communication imperfections.
  • Gradient-Free Optimization (Powell's Method): To achieve adaptive locomotion, the parameters of the CPGs are optimized online. The authors employ Powell's method, a gradient-free optimization algorithm, to search for the optimal set of CPG parameters. The robot's overall velocity is used as the fitness criterion for the optimization. This approach allows for parallel execution of the CPG and optimization algorithm, enabling real-time learning while the robot is moving.

Additionally, the paper details the YaMoR platform itself, including its mechanical and electronic design, and introduces the Scatternet Protocol (SNP), a Bluetooth-based communication system extended to support more than 8 nodes by forming a scatternet. This protocol is crucial for enabling the distributed CPG control across multiple modules.

Datasets

Not discussed in the paper. The experiments were conducted on the developed YaMoR hardware platform.

Key Insights and Ideas

  • CPGs as Robust Building Blocks for Modular Robots: The paper highlights that CPGs are well-suited for modular self-reconfigurable (MSR) robots due to their inherent robustness to communication noise and their ability to generate coordinated rhythmic patterns.
  • Online Learning for Real-Time Adaptation: The combination of CPGs with Powell's optimization method enables a truly online learning strategy. This means the robot can adapt its locomotion patterns while in motion, without needing to stop or rely on pre-computed models.
  • Gradient-Free Optimization's Efficiency: Powell's method is shown to be effective with a relatively small number of fitness evaluations, which is critical for real-time learning on hardware with limited battery life and processing power.
  • Scatternet Protocol for Scalable Communication: The SNP is a practical solution for enabling communication in larger modular robot systems, overcoming the limitations of standard Bluetooth piconets.

Key Results

The paper demonstrates successful application of the proposed adaptive locomotion control on various modular robot configurations, including snake, turtle, tripod, and quadruped.

  • For the tripod configuration, good gaits (achieving speeds of 5 cm/s or better) were found within approximately 15 minutes of optimization.
  • The optimization results showed convergence, indicating that the algorithm found a stable set of parameters for efficient locomotion.
  • The system proved robust to communication imperfections, with the CPGs maintaining stable oscillations and synchronization.

Limitations and Future Work

  • Limited Number of Evaluations: While Powell's method is efficient, the authors aim to further decrease the number of evaluations required to enable even faster adaptation to environmental or topological changes.
  • Noise in Velocity Measurement: The nonlinear displacement of the center module in some configurations introduced noise in the velocity measurement, which could potentially hinder the optimization from finding a "narrow" global optimum. However, this noise might also be beneficial in helping the algorithm escape local optima.
  • Autonomous Reconfiguration: The current reconfiguration of modular robots is manual. The authors are developing a centralized reconfiguration strategy based on graph signatures to enable autonomous changes in robot topology.
  • Changing Topologies and Environments: The future work explicitly mentions extending the adaptive locomotion learning to changing modular robot topologies and environmental situations.

Summary updated

Abstract

BIOROB
BibTeX
@inproceedings{sproewitz2007adaptive,
  author = {Alexander Sproewitz and Rico Moeckel and Jérôme Maye and Masoud Asadpour and Auke Jan Ijspeert},
  title = {{Adaptive Locomotion Control in Modular Robotics}},
  booktitle = {Infoscience},
  year = {2007},
}
Figure