ZSL-RPPO: Zero-Shot Learning for Quadrupedal Locomotion in Challenging Terrains using Recurrent Proximal Policy Optimization

TL;DR

ZSL-RPPO addresses the challenge of robust quadrupedal locomotion on challenging terrains without real-world fine-tuning by replacing teacher-student imitation with a recurrent PPO-based framework. The authors introduce RPPO for partially observable settings and a GRPN architecture to fuse proprioceptive and exteroceptive data, trained with extensive domain randomization. They demonstrate zero-shot transfer to Unitree A1 and Aliengo using Lidar or depth sensing across stairs, oily surfaces, sand, and cobblestones, outperforming state-of-the-art baselines. This work reduces the reliance on costly real-world trials and offers a scalable approach for robust legged locomotion in diverse environments.

Abstract

We present ZSL-RPPO, an improved zero-shot learning architecture that overcomes the limitations of teacher-student neural networks and enables generating robust, reliable, and versatile locomotion for quadrupedal robots in challenging terrains. We propose a new algorithm RPPO (Recurrent Proximal Policy Optimization) that directly trains recurrent neural network in partially observable environments and results in more robust training using domain randomization. Our locomotion controller supports extensive perturbation across simulation-to-reality transfer for both intrinsic and extrinsic physical parameters without further fine-tuning. This can avoid the significant decline of student's performance during simulation-to-reality transfer and therefore enhance the robustness and generalization of the locomotion controller. We deployed our controller on the Unitree A1 and Aliengo robots in real environment and exteroceptive perception is provided by either a solid-state Lidar or a depth camera. Our locomotion controller was tested in various challenging terrains like slippery surfaces, Grassy Terrain, and stairs. Our experiment results and comparison show that our approach significantly outperforms the state-of-the-art.

Figures (8)

• Figure 1: Overview of the locomotion control pipeline.
• Figure 2: Policy network architecture.
• Figure 3: We narrow the perception gap between simulation and the real world by applying pre-processing techniques and 3D reconstruction.
• Figure 4: Behaviour-tuning rewards: the top two subplots illustrate the effects of the "foot stance" reward, which encourages the quadruped to place its footholds away from the edges of stair treads; the bottom two subplots illustrate the "stumble" reward effect, which prevents stumbled legs.
• Figure 5: Selected terrains for real-world evaluation.