Quantifying canine interactions with smart toys assesses suitability for service dog work

Abstract

There are approximately a half million active service dogs in the United States, providing life-changing assistance and independence to people with a wide range of disabilities. The tremendous value of service dogs creates significant demand, which service dog providers struggle to meet. Breeding, raising, and training service dogs is an expensive, time-consuming endeavor which is exacerbated by expending resources on dogs who ultimately will prove to be unsuitable for service dog work because of temperament issues. Quantifying behavior and temperament through sensor-instrumented dog toys can provide a way to predict which dogs will be suitable for service dog work, allowing resources to be focused on the dogs likely to succeed. In a 2-year study, we tested dogs in advanced training at Canine Companions for Independence with instrumented toys, and we discovered that a measure of average bite duration is significantly correlated with a dog's placement success as a service dog [Adjusted OR = 0.12, Pr(>|z|) = 0.00666]. Applying instrumented toy interactions to current behavioral assessments could yield more accurate measures for predicting successful placement of service dogs while reducing the workload of the trainers.

Keywords: Canine Companions for Independence; animal behavior; computational behavior; instrumented toys; quantified interactions.

Figure 1

Ball-Human Sensor experiment setup. The experimenter rolls the ball, and then (left) the dog is mildly restrained for one second, and (middle) released to pursue the ball. The dog then retrieves and interacts with the ball (right).

Figure 2

Ball-Ramp experiment setup. The experimenter drops the ball down the enclosed ramp and then (left) the dog is mildly restrained for one second, and (middle) released to pursue the ball. The dog then retrieves and interacts with the ball (right).

Figure 3

Ball sensor. left: outer ball and inner ball; right: electronics and battery that are placed inside the inner ball.

Figure 4

A visual description of the engineered bite.

Figure 5

Correlation matrix of CCI features collected by the instrumented ball (modeling feature importance with respect to class).

Figure 6

Likelihood and distribution plots for each feature. (A) Top: The likelihood of passing the CCI criteria using average bite strength plotted by condition. Bottom: Distribution of average bite strength for each condition plotted against CCI outcome. (B) Top: The likelihood of passing the CCI criteria using peak bite frequency plotted by condition. Bottom: Distribution of peak bite frequency for each condition plotted against CCI outcome. (C) Top: The likelihood of passing the CCI criteria using average bite duration plotted by condition. Bottom: Distribution of average bite duration for each condition plotted against CCI outcome. (D) Top: The likelihood of passing the CCI criteria using interaction time plotted by condition. Bottom: Distribution of interaction time for each condition plotted against CCI outcome.