Automatic Snake Gait Generation Using Model Predictive Control

Emily Hannigan, Bing Song, Gagan Khandate, Maximilian Haas-Heger, Ji Yin, Matei T. CiocarlieICRA 2020 · DOI

TL;DR

This paper introduces a novel Model Predictive Control (MPC) approach for automatically generating snake robot gaits without relying on predefined patterns. By optimizing trajectories based on environmental models, the method produces efficient and adaptable gaits that match or exceed human-tuned serpenoid curves, even enabling complex maneuvers like sharp turns and obstacle avoidance.

What Problem Does This Paper Address?

Snake robots, with their high degrees of freedom, offer great potential for navigating challenging and cluttered environments. However, effectively controlling their locomotion is a significant challenge. Traditional methods often rely on predefined gait patterns, like the serpenoid curve, which require extensive manual tuning for each specific environment (e.g., dry friction, viscous water, fluid dynamics). This manual tuning is tedious, and the resulting gaits rarely transfer well between different conditions. The core problem this paper addresses is the need for an automatic, adaptive, and generalizable method for generating efficient snake robot gaits that can perform well across diverse environments without requiring re-tuning.

Methods

The authors propose using Model Predictive Control (MPC) for automatic gait generation. Instead of starting with a fixed movement pattern, MPC optimizes a sequence of control actions (joint torques) over a finite horizon to minimize a cost function. The key innovation lies in integrating a detailed dynamic model of the snake robot with various environmental reaction force models into the MPC framework.

  • Model Predictive Control (MPC): The core algorithm used is an iterative Linear Quadratic Regulator (iLQR), a variant of MPC. It repeatedly computes an optimal control sequence based on the current state and a predictive model of the system's dynamics.
  • Dynamic Model: A 2D n-link kinematic chain model is employed, incorporating Newtonian dynamics. This model is crucial for predicting how the robot will move in response to control inputs.
  • Environmental Reaction Force Models: The paper introduces and utilizes sophisticated models for different environments:
  • Anisotropic Coulomb Friction: A novel, smooth model is presented that obeys the Maximum Dissipation Principle, capturing anisotropic dry friction for overland locomotion and allowing for side slip.
  • Viscous Friction: Models anisotropic viscous friction, relevant for low Reynolds number underwater locomotion.
  • Drag and Added Mass: Accounts for fluid dynamics effects at high Reynolds numbers, crucial for underwater movement.
  • Cost Function: The cost function guides the MPC optimization. For straight-line locomotion, it prioritizes reaching a goal position and minimizing power consumption. For obstacle avoidance, an additional term penalizes proximity to obstacles.

Datasets

Not discussed in the paper. The paper focuses on simulation-based evaluation using synthesized gaits.

Key Insights and Ideas

  • Environment Generalizability: The most striking insight is that a single set of MPC parameters can generate effective gaits across vastly different environments (dry friction, viscous fluid, high-Reynolds number fluid) without any re-tuning. This highlights the power of model-based optimization adapting to environmental physics.
  • Beyond Serpenoid: The MPC approach is not limited to serpenoid-like gaits. By modifying the cost function, it can generate more complex and task-specific movements, such as sharp turns or navigating narrow tunnels while avoiding contact.
  • Anisotropic Friction Model: The introduction of a smooth, physically grounded anisotropic Coulomb friction model is a significant contribution for more realistic overland snake robot locomotion.
  • Qualitative Mirroring of Nature: The generated gaits, particularly in the fluid dynamics environment, exhibit amplitude variations in joint movements that qualitatively resemble natural snake and eel locomotion.

Key Results

  • Efficiency: MPC-generated gaits are comparable to, and in some cases better than, Pareto-optimal serpenoid gaits in terms of speed and power consumption. For instance, in dry friction, MPC achieved higher speeds at similar power levels compared to tuned serpenoid gaits.
  • Adaptability: A single set of MPC parameters (α=1, β=0.01, xg=-20) successfully generated efficient gaits across all tested environments, demonstrating remarkable generalizability.
  • Complex Gaits: The paper successfully demonstrates the generation of gaits for sharp turns and obstacle avoidance through tunnels, showcasing the flexibility of the MPC approach beyond simple forward locomotion.
  • Computational Performance: For a 6-second gait, the MPC optimization took between 1.0 and 1.4 seconds on a commodity processor, indicating that while promising, further speed-up is needed for real-time robotic applications.

Limitations and Future Work

  • Computational Cost: The primary limitation acknowledged is the computational cost of the iLQR algorithm. While the current implementation is not fast enough for real-time execution, the authors estimate a 10X-20X speed-up is achievable through optimized implementations and potentially more extensive parallelization.
  • Model Accuracy: The performance of MPC is inherently tied to the accuracy of the dynamic model. The paper mentions preliminary tests on robustness to errors in viscous friction coefficients, showing minor performance degradation with small errors. A more comprehensive evaluation of robustness to various model parameter errors is needed.
  • Real-World Deployment: Bridging the gap between simulation and real-world deployment requires addressing the computational cost and ensuring the dynamic model accurately reflects real-world physics.
  • Future Work: The authors suggest exploring algorithmic improvements for speed-up, investigating parallelization, and conducting more thorough robustness analyses. They also envision this MPC approach complementing Central Pattern Generator (CPG) methods for creating more versatile robot snakes capable of operating in complex, unstructured environments.

Summary updated

BibTeX
@inproceedings{hannigan2020automatic,
  author = {Emily Hannigan and Bing Song and Gagan Khandate and Maximilian Haas-Heger and Ji Yin and Matei T. Ciocarlie},
  title = {{Automatic Snake Gait Generation Using Model Predictive Control}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2020},
  doi = {10.1109/icra40945.2020.9196853},
}
Figure