Integrated Sensing and Control Algorithms for Computer-Assisted Training
Abstract:
The aims of this project are to contribute the fundamental physical and algorithmic building blocks of a novel cyber-‐physical two-‐way communication platform designed to enable accurate training and monitoring of canines. The project efforts lie at the intersection of computer science, electrical engineering, and veterinary behavior. We report on our efforts in a number of areas including the design of new miniaturized sensors, the creation of new control algorithms inspired by human dog trainers, and empirical validation with both working dogs and untrained stock dogs.
The main contributions of the project can be grouped into three categories:
1. Hardware Development: the creation of novel sensor and actuator packages that enable real time, wireless monitoring of and communication with canines
2. Software Development: research and development of novel machine learning algorithms to support interpretation of dogs' behaviors and emotions using data collected from sensors, as well as computer-‐assisted reinforcement of behaviors for automated training applications
3. Validation: conducting novel experiments to demonstrate CPS capabilities The first year of the project has been very busy, and we have made exciting progress towards all three of our goals.
We have completed a prototype of the wearable physiological sensors, and further completed their preliminary testing. The physiology strap contains electrocardiography (ECG) electrodes using commercially-‐available off the shelf (COTS) dog electronic collar leads. The strap also has a Photoplethysmogram (PPG) sensor that uses light pipes to measure tissue diffusion. Both the ECG and PPG sensor use custom circuitry we designed to enable a wearable form factor. The data collected using our sensors has been validated against COTS monitors, and is accurate without shaving the dog’s fur. We have successfully completed our first phase of computer-‐assisted training, whereby a computer has trained two dogs to sit to get food rewards from a computer-‐operated treat dispenser. The training experiments leveraged the smart harness, posture recognition algorithms, and remotely operated treat dispensers to provide computer-‐delivered rewards to dogs. Our efforts involved exploring the tradeoff between classification accuracy and response latency. When we achieved response latency below 0.2s (which occurred in two of seven trials) dogs successfully learned to sit to get treats from dispensers without any influence from human handlers. Additionally, we spent six months contributing to the SmartAmerica Challenge, sponsored by the Whitehouse Office of Science and Technology Policy. The SmartAmerica Challenge put us together with four other Universities, and four companies, to demonstrate how Cyber-‐Physical Systems can be used to improve emergency response to save lives. Our work as part of the Smart Emergency Response Systems (SERS) team emphasized connected canines, and their role in search in rescue operations of the future. Through participation with the eight other partner organizations we matured our smart harness to facilitate interaction with a range of other technologies. Rapid prototype and testing as part of an integrated SERS testbed contributed to improved haptic feedback functionality, augmented environmental monitoring capabilities, and a more robust and reliable canine body-‐area network platform. As part of our participation in the SmartAmerica Challenge, our team presented at the "expo" which took place at the convention center in Washington, D.C. The Expo was extremely well-‐attended by government officials, reporters, and the general public. It led to numerous news reports on the technology.