Abstract

Achieving campus autonomy with robot dogs is a significant challenge that requires the robots to navigate various terrains and environments. Existing methods either focus on specialized skills or lack vision-based capabilities. In this work, we plan to propose a system for learning vision- based autonomy skills without relying on reference animal data or complex reward systems. We will utilize a reinforcement learning approach inspired by direct collocation to enable the robot dogs to autonomously navigate campus terrains, including climbing over obstacles, traversing uneven surfaces, crossing gaps, and adapting to different terrain types. Our system utilizes the robot dogs' onboard sensors, such as depth cameras, to perceive and react to the environment. Through experiments and evaluations, we demonstrate the effectiveness of our approach in enabling robot dogs to autonomously navigate and achieve campus autonomy over various terrains.

Introduction

Humans and animals possess amazing athletic intelligence. Parkour is an examplar of athletic intelligence of many biological beings capable of moving swiftly and overcoming various obstacles in complex environments by running, climbing, and jumping. Such agile and dynamic movements require real- time visual perception and memorization of surrounding environments, tight coupling of perception and action, and powerful limbs to negotiate barriers. One of the grand challenges of robot locomotion is building autonomous parkour systems. Boston Dynamics Atlas robots have demonstrated stunning parkour skills. However, the massive engineering efforts needed for modeling the robot and its surrounding environments for predictive control and the high hardware cost prevent people from reproducing parkour behaviors given a reasonable budget. Recently, learning-based methods have shown robust performance on walking, climbing stairs, mimicking animals and legged mobile manipulation by learning a policy in simulation and transferring it to the real world while avoiding much costly engineering and design needed for robot-specific modeling.

We introduces a robot parkour learning system for low-cost quadrupedal robots that can perform various parkour skills, such as climbing over high obstacles, leaping over large gaps, crawling beneath low barriers, squeezing through thin slits, and running. Our reinforcement learning method is inspired by direct collocation and consists of two simulated training stages: RL pre-training with soft dynamics constraints and RL fine-tuning with hard dynamics constraints. In the RL pre-training stage, we allow robots to penetrate obstacles using an automatic curriculum that enforces soft dynamics constraints. This encourages robots to gradually learn to overcome these obstacles while minimizing penetrations. In the RL fine-tuning stage, we enforce all dynamics constraints and fine-tune the behaviors learned in the pre-training stage with realistic dynamics. In both stages, we only use a simple reward function that motivates robots to move forward while conserving mechanical energy. After each individual parkour skill is learned, we use DAggerto distill them into a single vision-based parkour policy that can be deployed to a legged robot using only onboard perception and computation power.

Experiment Results

We first trained our models in simulation using Isaac Gym. Then we deploy our terrain-over policy on the Unitree Go2 robot dog with 12 leg joints. We use the Intel RealSense D435i to get the depth image in real time at frequency of 30Hz. Our robot now can easily and stably climb onto a 35cm high box and even a 45cm higher one sometimes. This is much higher than each stair we may encounter in the real campus environment, so our robot can easily go over such terrains, going up and down stairs.