Coordinating Spinal and Limb Dynamics for Enhanced Sprawling Robot Mobility

TL;DR

This work tackles robust, adaptable crawling for a salamander-inspired quadruped by hybridizing biologically grounded Hildebrand gaits with deep reinforcement learning, evaluating fixed and active spinal joints. Through MuJoCo simulations and real-world tests, the hybrid approach improves robustness on rough terrain and offers faster real-world traversal than Hildebrand alone, while pure end-to-end RL struggles with physical constraints. The study highlights the value of integrating principled gait priors with learning-based adaptation to achieve efficient, resilient spinal actuation in sprawled locomotion and discusses practical steps to bridge the sim-to-real gap. Future directions include formal reward shaping with STL and biologically inspired priors like Central Pattern Generators to further improve safety and stability in real-world operation.

Abstract

Sprawling locomotion in vertebrates, particularly salamanders, demonstrates how body undulation and spinal mobility enhance stability, maneuverability, and adaptability across complex terrains. While prior work has separately explored biologically inspired gait design or deep reinforcement learning (DRL), these approaches face inherent limitations: open-loop gait designs often lack adaptability to unforeseen terrain variations, whereas end-to-end DRL methods are data-hungry and prone to unstable behaviors when transferring from simulation to real robots. We propose a hybrid control framework that integrates Hildebrand's biologically grounded gait design with DRL, enabling a salamander-inspired quadruped robot to exploit active spinal joints for robust crawling motion. Our evaluation across multiple robot configurations in target-directed navigation tasks reveals that this hybrid approach systematically improves robustness under environmental uncertainties such as surface irregularities. By bridging structured gait design with learning-based methodology, our work highlights the promise of interdisciplinary control strategies for developing efficient, resilient, and biologically informed spinal actuation in robotic systems.

Figures (8)

• Figure 1: Biological and robotic model. (left) The fire salamander (Salamandra salamandra) is a common species of salamander found in Europe. (middle) The robot model in the Mujoco simulation world. (right) 3D printed robotic salamander.
• Figure 2: Simulation Environment and Robot Actuation. (Left) The MuJoCo simulation environment used for evaluating the robot's locomotion. The red sphere indicates the target goal. (Right) CAD model of the robot detailing its kinematic structure. The red arrows represent the two rotational axes per leg, and the black arrow indicates the single spinal joint that enables lateral bending for sprawling locomotion.
• Figure 3: Example biological gait-cycle. Joint positions during one walking cycle following the Hildebrand-style gait for both the passive and active spinal joint scenarios. Each leg spends 25% of the cycle in the air and 75% on the ground.
• Figure 4: State variables and Results. (Left) Observation space for the 8-joint and 9-joint robot configurations. (Right) RL Learning Curves.
• Figure 5: Reachability with and without an Active Spinal Joint. The figure compares the forward reachability of the robot with a flexible spine (9-joints) and a rigid spine (8-joints). As shown, the active spinal joint enables the robot to extend its reach farther forward ($L_1 > L_2$) for the same leg joint angles.