Bridge the Gap: Enhancing Quadruped Locomotion with Vertical Ground Perturbations

TL;DR

This work addresses the robustness of quadruped locomotion to vertical ground perturbations by evaluating policies trained on an oscillating bridge. It employs PPO-based reinforcement learning in MuJoCo with $48$ parallel environments and broad domain randomization to achieve zero-shot transfer to a real Unitree Go2 walking on a $13.24$ m oscillating bridge with eigenfrequency $2$ Hz. Fifteen policies across five gaits and three training conditions are compared, showing that exposure to vertical perturbations during training yields superior stability over policies trained on rigid ground. The results demonstrate successful sim-to-real transfer and provide guidelines for designing gait-robust quadrupeds operating on vibrating or unstable terrains.

Abstract

Legged robots, particularly quadrupeds, excel at navigating rough terrains, yet their performance under vertical ground perturbations, such as those from oscillating surfaces, remains underexplored. This study introduces a novel approach to enhance quadruped locomotion robustness by training the Unitree Go2 robot on an oscillating bridge - a 13.24-meter steel-and-concrete structure with a 2.0 Hz eigenfrequency designed to perturb locomotion. Using Reinforcement Learning (RL) with the Proximal Policy Optimization (PPO) algorithm in a MuJoCo simulation, we trained 15 distinct locomotion policies, combining five gaits (trot, pace, bound, free, default) with three training conditions: rigid bridge and two oscillating bridge setups with differing height regulation strategies (relative to bridge surface or ground). Domain randomization ensured zero-shot transfer to the real-world bridge. Our results demonstrate that policies trained on the oscillating bridge exhibit superior stability and adaptability compared to those trained on rigid surfaces. Our framework enables robust gait patterns even without prior bridge exposure. These findings highlight the potential of simulation-based RL to improve quadruped locomotion during dynamic ground perturbations, offering insights for designing robots capable of traversing vibrating environments.

Figures (7)

• Figure 1: We train the Unitree Go2 quadruped in simulation and zero-shot transfer the learned policies to the oscillating real-world humvib bridge to investigate the impact of vertical ground perturbations on locomotion robustness.
• Figure 2: Schematic of the Unitree Go2 quadruped crossing the 13.24-meter long humvib bridge.
• Figure 3: Top: Episode return for the different gaits (left) and training conditions (right) during training. Bottom: Episode return over the command velocity on the oscillating bridge for the different gaits (left) and training conditions (right) of the final policies during evaluation.
• Figure 4: Footfall pattern of the different gait styles on the rigid bridge with a target speed $v_x = 0.5$ m/s. The feet of the robot are denoted with front left (FL), front right (FR), rear left (RL), and rear right (RR). In our study, we omit the pronk gait, as we were unable to train a policy that sufficiently learned the desired footfall pattern, due to the agent resorting to a more stable trot-like gait during learning.
• Figure 5: Movement of the robot's with a target speed $v_x = 0.5$ m/s for the default gait pattern under three conditions: no oscillation (nos), equidistant bridge (eb), and equidistant ground (eg). Subfigures on the first column show x, y, and z directions of robot's on an idle bridge, while the subfigures on the second column show the same quantities, but on an oscillating bridge.