Adaptive and Energy-Efficient Optimal Control in CPGs Through Tegotae-Based Feedback

Riccardo Zamboni, Dai Owaki, Mitsuhiro HayashibeFrontiers Robotics AI 2021 · OpenReview

TL;DR

This paper introduces a novel "Tegotae-based" feedback control strategy for Central Pattern Generators (CPGs) that aims to achieve both adaptive locomotion and optimal energy efficiency in robots. By integrating proprioceptive feedback with the CPG's internal state and employing a reflex-like actuation, the approach demonstrates improved adaptability to environmental changes and comparable energy efficiency to optimal control methods.

What Problem Does This Paper Address?

Efficient and adaptive locomotion is a fundamental challenge in robotics, mimicking the sophisticated capabilities of biological systems. Central Pattern Generators (CPGs) are neural circuits that naturally produce rhythmic motor patterns, making them a popular choice for robotic locomotion control. While CPGs can generate basic rhythms without sensory input, incorporating sensory feedback is crucial for adapting to dynamic environments. However, adding feedback often complicates CPG dynamics, leading to unpredictable behavior like bifurcations. This paper tackles the problem of designing a systematic and effective sensory feedback mechanism for CPGs that enhances adaptability and optimizes energy consumption, moving beyond traditional approaches that often require extensive tuning or complex models.

Methods

The core of this work is the "Tegotae approach," a concept that quantifies the "goodness" of a perceived reaction relative to the controller's intention. This is mathematically formulated as a Tegotae function, $T(u, e)$, which is a product of the controller's intention ($C(u)$) and the sensory feedback ($S(e)$). The feedback signal ($f$) is then derived as the gradient of this Tegotae function with respect to the control variable ($u$), aiming to maximize $T$.

For CPGs, the Tegotae feedback is integrated into the phase oscillator dynamics: $\phi_i = \omega_i + f_i(\phi_i, e)$. A key innovation is the proposed "reflex-like actuation" ($F_a$), which is designed to be opposite in sign to the Tegotae feedback and becomes active only when the feedback indicates a "bad" Tegotae (i.e., a deviation from the intended action). This reflex-like actuation is defined as $F_a(\phi_i, F_k) = -\min(0, f_i)$, where $f_i$ is the Tegotae feedback. This approach aims to provide a direct, biologically inspired mechanism for correcting deviations and improving performance.

The authors validate their approach using two mechanical models: a monopod (single-legged hopper) and a biped (two-legged hopper). These models incorporate CPGs (modeled as Kuramoto oscillators) coupled with spring-damper systems and actuators. Simulations are performed to evaluate adaptation to environmental changes (e.g., ground level shifts) and energy efficiency, comparing the Tegotae approach with other feedback methods and an optimal control solution.

Datasets

Not discussed in the paper. The evaluation is based on simulations of custom-built mechanical models.

Key Insights and Ideas

  • Tegotae as a unifying concept: The Tegotae framework provides a principled way to design sensory feedback by focusing on the consistency between intended and perceived actions, offering a more intuitive approach than purely optimizing performance metrics.
  • Reflex-like actuation for efficiency: The proposed reflex-like actuation, directly coupled to the Tegotae feedback, allows for rapid, localized corrections, contributing to both adaptability and energy efficiency by only acting when necessary.
  • Embodied intelligence in action: The Tegotae approach embodies the principle of "embodied intelligence," where the physical body and its interaction with the environment play a crucial role in shaping control strategies.
  • Energy efficiency comparable to optimal control: The study demonstrates that the Tegotae approach can achieve energy efficiency comparable to explicitly designed optimal controllers, suggesting that biologically inspired methods can be highly performant.
  • Tacit learning parallels: The Tegotae approach exhibits similarities to "tacit learning" frameworks, where adaptation occurs through accumulating error signals, hinting at a deeper connection between biological control and machine learning.

Key Results

  • Monopod Hopping: The Tegotae feedback ($f_1$) demonstrated superior performance in terms of synchronization transient period and energy efficiency compared to other feedback types (height feedback $f_2$, Buchli's feedback $f_3$, and force feedback $f_4$). For instance, the energy efficiency ($E_e$) for Tegotae was 1.50 m/Ws, while other methods ranged from 1.15 to 1.25 m/Ws, with $f_4$ failing to provide a stable orbit.
  • Adaptivity: The Tegotae approach showed robust adaptation to environmental changes, such as a sudden drop in ground level, by re-polarizing the oscillator and adjusting force injection without explicit parameter tuning.
  • Bipedal Gait Transition: The Tegotae control policy successfully achieved distinct in-phase and anti-phase bipedal hopping gaits by simply changing the oscillator frequency ($\omega$), without requiring explicit inter-oscillator coupling.
  • Optimal Control Comparison: Simulations comparing the Tegotae controller with an optimal controller (using multiple shooting methods) showed remarkably similar control actions, position, and velocity profiles, and comparable energy efficiency across different masses. For example, in MS1 (m=0.1), the Tegotae controller achieved an RMSE of 0.22, similar to the optimal controller.

Limitations and Future Work

  • Real-world robot validation: The authors acknowledge that the applicability of the Tegotae approach to a physical robot with more degrees of freedom has not yet been tested.
  • Modeling complexity: Accurately modeling the dynamics of real-world environments remains a significant challenge.
  • Parameter tuning: While the Tegotae approach aims to reduce parameter tuning, some parameters (like $\sigma$ and $\Delta$) still require careful selection, and their effect on stability was noted but not fully explored.
  • Future Directions: The authors suggest applying the control policy to robots with higher degrees of freedom and more complex tasks. They also propose exploring the force/torque-based control of physical robots, which is compatible with their approach, as a promising avenue for future research. Further investigation into the optimal selection of Tegotae function parameters and their impact on adaptability and efficiency is also warranted.

Summary updated

Abstract

To obtain biologically inspired robotic control, the architecture of central pattern generators (CPGs) has been extensively adopted to generate periodic patterns for locomotor control. This is attributed to the interesting properties of nonlinear oscillators. Although sensory feedback in CPGs is not necessary for the generation of patterns, it plays a central role in guaranteeing adaptivity to environmental conditions. Nonetheless, its inclusion significantly modifies the dynamics of the CPG architecture, which often leads to bifurcations. For instance, the force feedback can be exploited to derive information regarding the state of the system. In particular, the Tegotae approach can be adopted by coupling proprioceptive information with the state of the oscillation itself in the CPG model. This paper discusses this policy with respect to other types of feedback; it provides higher adaptivity and an optimal energy efficiency for reflex-like actuation. We believe this is the first attempt to analyse the optimal energy efficiency along with the adaptivity of the Tegotae approach.
BibTeX
@inproceedings{zamboni2021adaptive,
  author = {Riccardo Zamboni and Dai Owaki and Mitsuhiro Hayashibe},
  title = {{Adaptive and Energy-Efficient Optimal Control in CPGs Through Tegotae-Based Feedback}},
  booktitle = {Frontiers Robotics AI},
  year = {2021},
  url = {https://openreview.net/forum?id=yHq7CK57R9},
}
Figure