标题：Learning to Walk by Steering: Perceptive Quadrupedal Locomotion in Dynamic Environments
Authors: Mingyo Seo, Ryan Gupta, Yifeng Zhu, Alexy Skoutnev, Luis Sentis, and Yuke Zhu
来源：2023 IEEE International Conference on Robotics and Automation (ICRA 2023)
This is the third article shared by Jiajiaguai
Summary
We want to solve sensorimotor problems in dynamic environments. In this problem, a quadruped robot must exhibit robust and agile walking behavior against environmental clutter and mobility obstacles. We propose a hierarchical learning framework named PRELUDE, which decomposes the perceptual-motor problem into high-level decisions to predict navigation instructions and low-level gait generation to achieve goal instructions. In this framework, we utilize human demonstrations collected on a steerable car to train a high-level navigation controller through imitation learning and a low-level gait controller using reinforcement learning (RL). Therefore, our method can capture complex navigation behaviors from human supervision and discover variable gaits from trial and error. We show that we demonstrate the effectiveness of our approach in simulations and hardware experiments. The effectiveness of our method is demonstrated in experiments.
Figure 1: Method overview. PRELUDE solves perceptual motion problems in dynamic environments. We introduce a control hierarchy in which a high-level controller is trained through imitation learning to set navigation instructions, and a low-level gait controller is trained through reinforcement learning to achieve target instructions through joint spatial execution. This combination enables us to efficiently deploy the entire hierarchy on a quadruped robot in a real-world environment.
Figure 2: Model structure of PRELUDE. The advanced navigation strategy generates the target speed command ut at a frequency of 10Hz based on the observation data of the onboard RGB-D camera and the robot's heading. The target velocity command is used as input to the low-level gait controller together with the velocity command buffer Bt ut, the most recent robot state qt, and the previous joint space action at-1. The low-level gait strategy predicts joint space motions into required joint positions at 38Hz and sends them to the quadruped robot for actuation.
Figure 3: Hardware platform. A steerable cart designed to collect human navigation demonstrations (left) and a Unitree A1 robot with an egocentric RGB-D camera mounted at cart height (right). It ensures that navigation strategies trained on demonstration data can be deployed directly to the robot.
Figure 4: Real robot experiment. (Pictured left) We conducted real-world experiments with robots traversing a 15-meter track in different configurations. The figure shows the distribution of walking distance in meters. The black and red lines represent the range and average of the crossing lengths, respectively. (Right picture) We observed that PRELUDE (A1 default gait) drifted violently and hit the wall after turning at high speed, while PRELUDE quickly turned around the walking crowd and successfully completed the experiment.
Table I: Comparison of PRELUDE with baseline in simulations. We report the average traversal length in meters (total track length: 50 meters) and the success rate in percent as evaluation metrics.
Table
2: Gait controller evaluation We report tracking error in meters (smaller is better) and tracking success rate in percentage (higher is better).
in conclusion
We introduce PRELUDE as an efficient method for learning perceptual motion controllers for quadruped robots to traverse real-world dynamic environments. Our approach combines the complementary advantages of imitation learning and reinforcement learning to decompose the locomotion problem into high-level navigation and low-level gait generation through a hierarchical design. We designed a steerable car platform to collect demonstrations of human navigation in complex scenes and used the collected datasets to train advanced navigation strategies. We use large-scale reinforcement learning to train a low-level gait controller in simulations, demonstrating that it transfers effectively to the real world and produces robust and variable motion. Our work mainly focuses on flat ground and indoor environments where human steering motions can be conveniently collected by wheeled platforms. In future work, we hope to extend our wheeled cart into more complex mechanical designs to collect datasets on humans walking on rough terrain outdoors.