Learning Omnidirectional Locomotion for a Salamander-Like Quadruped Robot

TL;DR

This work addresses the limitation of predefined gait patterns for salamander-like quadrupeds by introducing a phase-based learning framework that enables omnidirectional locomotion without reference motions. Each body part is governed by a bidirectional phase variable and guided by a phase-coverage reward, while symmetry-based data augmentation enforces motion- and task-level symmetry to improve learning efficiency and gait quality. The approach yields a rich repertoire of 22 gaits, including forward, lateral, diagonal, and in-place rotations, with smooth transitions between gaits and robust performance in both simulation and real-world experiments. The results demonstrate the practical impact of combining phase-based control with morphology-aware data augmentation, offering a scalable path toward versatile bio-inspired robots and paving the way for perception-guided, multi-morphology locomotion.

Abstract

Salamander-like quadruped robots are designed inspired by the skeletal structure of their biological counterparts. However, existing controllers cannot fully exploit these morphological features and largely rely on predefined gait patterns or joint trajectories, which prevents the generation of diverse and flexible locomotion and limits their applicability in real-world scenarios. In this paper, we propose a learning framework that enables the robot to acquire a diverse repertoire of omnidirectional gaits without reference motions. Each body part is controlled by a phase variable capable of forward and backward evolution, with a phase coverage reward to promote the exploration of the leg phase space. Additionally, morphological symmetry of the robot is incorporated via data augmentation, improving sample efficiency and enforcing both motion-level and task-level symmetry in learned behaviors. Extensive experiments show that the robot successfully acquires 22 omnidirectional gaits exhibiting both dynamic and symmetric movements, demonstrating the effectiveness of the proposed learning framework.

Figures (8)

• Figure 1: Overview of the control framework, including the robot prototype and the body-fixed coordinate system. The robot has 15 actuated degrees of freedom in total, consisting of a segmented spine with 3 DOFs that allows horizontal bending and four legs, each with 3 joints. The origin of the body coordinate system is located at the midpoint between the front and hind girdles, and the x-axis points from the hind girdle toward the front girdle.
• Figure 2: Mapping from phase variables to leg joint angles. Colors represent different phase values, where the leg is in the swing phase for $[0,\pi)$ and in the stance phase for $[\pi,2\pi)$.
• Figure 3: Diagram of the morphological symmetries of the salamander-like quadruped robot. Colors indicate phase values consistent with those in Fig. \ref{['fig:tg']}.
• Figure 4: Schematic of the experimental platform.
• Figure 5: Simulated robot trajectories under seven gait categories (see Table. \ref{['tab:gait_type']}). Each trajectory is visualized as a sequence of arrows, with arrow orientation indicating body yaw and arrow color fading from light to dark to represent temporal progression. The rotational gaits are shown in the second and third subplots, corresponding to in-place counterclockwise and clockwise rotations.