Abstract
Legged robots excel on static, uneven terrain but often fail when the ground itself is in motion. This study investigates quadruped locomotion on surfaces undergoing vertical oscillations, specifically focusing on a 13.24-meter bridge at its resonant frequency. Using Deep Reinforcement Learning (DRL) with Proximal Policy Optimization (PPO), we trained a Unitree Go2 robot under various height-regulation strategies and gait constraints. Our results demonstrate that incorporating vertical perturbations into the training loop and rewarding surface-relative height regulation enables robust zero-shot transfer from simulation to a physical vibrating bridge. The proposed "Equidistant Bridge" (eb) strategy effectively mitigates lateral drift and reduces foot impact forces compared to traditional rigid-ground controllers.
Introduction
While modern quadrupeds navigate stairs and rocky slopes with ease, active ground movement remains a significant challenge. Industrial platforms, disaster zones, and suspension bridges often exhibit vertical vibrations that can destabilize traditional controllers assuming a static base. This research addresses this "blind spot" by exploring how different training strategies and gait patterns influence a robot's stability on an oscillating surface.
Methods
The experimental setup utilized the HUMVIB bridge, a 13.24m steel-and-concrete structure with a 2.0 Hz eigenfrequency. In simulation (MuJoCo), the bridge was modeled as a harmonic oscillator. We employed PPO to train 15 distinct policies across three height-regulation conditions:
- No Oscillation (nos): Standard training on a rigid surface.
- Equidistant Bridge (eb): Height regulated relative to the moving bridge surface.
- Equidistant Ground (eg): Height regulated relative to the static ground beneath the bridge.
Training involved extensive domain randomization of robot mass, friction, and sensor noise to facilitate sim-to-real transfer. Five gait types (trot, pace, bound, free, and default) were evaluated.
Key Results
The "nos" policies, trained on rigid ground, performed poorly on the vibrating bridge, exhibiting significant lateral drift and chaotic behavior. In contrast, policies using the "eb" strategy successfully adapted by naturally "canceling out" perturbations through leg mechanics.
Key findings include:
- Zero-Shot Transfer: Policies transferred directly to the physical bridge without fine-tuning.
- Stability and Drift: On the real bridge (2.0 Hz, ±0.1m amplitude), "eb" and "eg" policies maintained straight paths, while "nos" policies drifted severely.
- Gait Emergence: Unconstrained "free" gaits naturally evolved into trot-like patterns, suggesting inherent stability in diagonal coordination for vertical oscillations.
- Force Mitigation: "eb" policies showed significantly lower foot impact forces, indicating the agent learned to "soften" its gait to maintain stability.
Discussion and Future Work
This work demonstrates that vertical ground perturbations can be effectively managed by integrating them into the RL training loop. However, limitations remain, such as the focus on a single resonant frequency and the failure of highly dynamic gaits like "pronk" to emerge as stable solutions. Future research will explore a broader spectrum of vibration frequencies and more complex, non-linear terrains, potentially integrating visual navigation for autonomous traversal of unstable environments.
Conclusion
By shifting the focus from static to dynamic ground conditions, this research provides a framework for enhancing the operational range of legged robots in real-world industrial and disaster-response scenarios.
Summary updated