Ground Contact and Reaction Force Sensing for Linear Policy Control of Quadruped Robot

Efforts have been made in the past years to reduce the cost of entry into legged robotics research, which has extended into development of computationally light-weight control policies that can be deployed on cost-effective hardware. While there is no restriction to the nature of inputs to such policies, they typically rely only on inertial data to observe the robot state. However, since legged robots primarily interact with the environment through ground contact and reaction forces, it is crucial to include these data in the observation space of a policy. This paper investigates the effect of including terrain sensing via ground contacts and ground reaction forces in the observation space on the ability of the policy to adapt robot behavior to terrain variations. Further, a comparison is made between the performance improvement when using ground contacts versus ground reaction forces. We present experimental data that demonstrate the advantages of including terrain sensing in combination with inertial data by improving robot survivability and stability. Lastly, based on the results of the comparison between the two methods of terrain sensing, we present a low-cost hardware implementation of ground contact detection.