Towards Safe and Stable Landing Control for Quadruped Robots

In the last decades, legged robots have increasingly matured and demonstrated versatile locomotion capabilities. Recent advances encourage the use of quadrupedal morphology to accomplish highly dynamic motions. Different types of gaits, such as trot or crawl, have been developed. However, sometimes there is no way to get around or over an obstacle with the gaits mentioned above, and jumping motions are required. In contrast to advances in jump planning, relatively little research has been done on landing strategies, that permit safely recovering after unexpected falls or planned jumps. These abilities are beneficial for navigating harsh environments and preventing significant damage to the robot. The overall research objective of my Ph.D. research aims to fill the gap on the topic of quadrupedal robots' safe and stable landing after high drops. Specifically, I decided to find a solution for the following two problems. 1. Can a quadruped robot have the aerial righting reflexes necessary to reorient itself as it falls? 2. Using only proprioceptive measurements, can a quadruped robot detect a high fall and react in order to safely land? To answer the first question, I observed that most quadrupeds are designed with lightweight legs, causing the limbs to have minimal impact on the overall angular momentum. Thus, I designed an Orientation Control System (OCS) composed of two rotating masses to gain control authority over the trunk orientation when the robot has no contact with the environment. The axes of rotation of the flywheels are set to be incident, enabling continuous controllability in both roll and pitch directions while keeping the device compact. The approach was tested in a simulation environment using the 2.5 kg torque-controlled quadruped robot Solo12. Not negligible horizontal velocity for the robot's Center of Mass (CoM) makes the second question much more challenging. In some situations, ignoring it is not an option. For instance, when a quadruped trots with a high speed and must do a leap without first stopping the motion. To tackle the landing problem, I proposed a reactive Landing Controller. Knowing neither the distance to the landing surface nor the flight time, it reacts to make the robot ready to dissipate the kinetic energy once the touch down is identified. Relying on the notions of capturability and Zero Moment Point, the proposed controller fulfills the following requirements: 1) no bounces should be produced after landing, 2) the trunk must not hit the ground, 3) the robot must reach a stable standing state, and 4) once landed, the feet must not slip. This is the first time a quadruped robot can successfully recover from falls with horizontal velocities up to 3 m/s in simulation. Experiments demonstrate that the torque-controlled quadruped Go1 can successfully achieve a stable configuration after falls when subjected to manually induced horizontal velocities and angular perturbations. When the horizontal velocity of the quadruped when it falls is larger than the maximum one the \gls{lc} can stabilize, reaching a stable standing state is impossible. The robot falls down and hardware damage can occur. Conscious of the robot's kinematic, actuation, and friction limits, the Extended Landing Controller (ELC) relaxes the above requirement. During the fall, it plans joint references that avoid collisions between the trunk and the ground and allow for one or more additional aerial phases, as the gymnasts do when they lose balance after touching the ground. Preliminary results considering Go1 free falling with a horizontal velocity of 4 m/s show that if reducing horizontal velocity between a touch down and the subsequent lift off is viable, the combined use of ELC and LC eventually leads the robot to a stable standing state. Locosim is the underlying enabling technology % my tutor Michele Focchi and I developed for these projects: it is an open-source, cross-platform robotics framework for working either in a simulation environment or with real hardware. Benefiting from Python and ROS and integrating features for computation of robots’ kinematics and dynamics, logging, plotting, and visualization, it considerably helps roboticists who need a starting point for rapid code prototyping.