PIE: Parkour with Implicit-Explicit Learning Framework for Legged Robots

TL;DR

This paper proposes a one-stage end-to-end learning-based parkour framework with Implicit-Explicit learning framework for legged robots (PIE) that leverages dual-level implicit-explicit estimation, and demonstrates successful zero-shot deployment of this framework on harsh terrains.

Abstract

Parkour presents a highly challenging task for legged robots, requiring them to traverse various terrains with agile and smooth locomotion. This necessitates comprehensive understanding of both the robot's own state and the surrounding terrain, despite the inherent unreliability of robot perception and actuation. Current state-of-the-art methods either rely on complex pre-trained high-level terrain reconstruction modules or limit the maximum potential of robot parkour to avoid failure due to inaccurate perception. In this paper, we propose a one-stage end-to-end learning-based parkour framework: Parkour with Implicit-Explicit learning framework for legged robots (PIE) that leverages dual-level implicit-explicit estimation. With this mechanism, even a low-cost quadruped robot equipped with an unreliable egocentric depth camera can achieve exceptional performance on challenging parkour terrains using a relatively simple training process and reward function. While the training process is conducted entirely in simulation, our real-world validation demonstrates successful zero-shot deployment of our framework, showcasing superior parkour performance on harsh terrains.

Figures (7)

• Figure 1: Our parkour framework with implicit-explicit estimation allows low-cost robots to traverse a series of challenging parkour-like terrains. Notably, in the real world the robot is not limited to the terrain on which it was trained in simulation, demonstrating impressive generalization capabilities.
• Figure 2: Overview of the proposed PIE framework. The grey box represents the estimator, which utilizes implicit-explicit estimation to provide the policy network with the estimated vectors. The framework is concurrently optimized with PPO for the actor and critic network and regression for the estimator.
• Figure 3: Simulation experiments results for PIE and PIE without $\hat{\mathbf{o}}_{t+1}$ in the presence of various camera input errors. Five plots represent the five types of camera errors introduced, with the x-axis of each plot representing: the distance deviates of the depth image from the actual terrain perception in the world coordinate system along the negative x-axis; the maximum value of uniformly distributed noise added to the position of the camera relative to the base link of the robot in the x, y and z directions; the maximum value of uniformly distributed noise added to the pitch of the camera coordinate system relative to the robot coordinate system; the standard deviation of Gaussian noise added to the depth image; the amount of salt-pepper noise added to the depth image, with random pixels being set to either the minimum or maximum depth based on the probability. These experiments adopt the highest difficulty level of terrains encountered by the robot during training, with the y-axis representing the average success rate across all terrains, including gap, stairs and step. Similarly, to reflect the generality of the experimental results, the average results of one hundred robots are taken for each experiment.
• Figure 4: Real-world indoor experiments results of each method. We measured the success rates for ten trails on each terrain of each difficulty level. Our PIE framework maintains consistent performance between real-world deployment and simulation, outperforming all other methods. Additionally, it demonstrates a certain level of generalization ability to traverse ramp terrains. Starred are previous related works zhuang2023robotcheng2023extreme.
• Figure 5: A 2km round-trip hike from the ZJU Yuquan campus to the top of Laohe Mountain and back, with an elevation gain of 153 meters from 27m to 180m. Our PIE framework successfully navigated the challenging terrains along this route, demonstrating notable generalization capabilities.