An intelligent mission-oriented and situation aware quadruped robot: a novel embodied explainable AI approach

Abstract

Quadruped locomotion and gait patterns as trotting, galloping, trot running has been investigated and applied to a variety of existing quadruped robots such as Big Dog from Boston Dynamics, A1 Robot dog from Unitree. Most of these studies are based either on biology inspired gaits or the best possible locomotion that can be performed by the individual robot with its pre-set mechanics and its availability of the degree of freedoms. While these are already available as their basic modes, a wide number of researchers are investigating locomotion via deep neural nets. These are making headlines in the research community for efficiency of use, and yet the explainability is lacking in most cases. Just like a Large Language Model might give spurious results here and there for basic common sense questions, these deep neural nets also make errors with unknown interpretability to the inputs. Regarding training, they require careful tuning of hyperparameters and training with a number of parameters unknown to user predictions. For example, on the field we might have a terrain which is flat for a certain length, in addition to a rocky climb, followed by a slippery slope. The combinations are as many as possible and the existing state of the art is heavily depending on human intervention and training predictions to handle the change of modes of the gait patterns that can fit into the terrain underneath. In this paper, we develop a novel embodied explainable machine learning algorithm which can help minimize the training as well as human intervention when autonomous operations are required. Specifically, we utilize the Markov Decision Process (MDP) along with rules set forth by DARPA in the Explainable AI (XAI) research. The XAI research enables us to generate textual explanations of the behavior by utilizing the MDP and reinforcement learning to generate mission oriented and situation aware cost functions along with the ones which are already pre-programmed. We validate our hypothesis in real hardware across different conditions.