
Teaching Robots Biologically Inspired Tasks Emma Van Hoogmoed B.S Candidate, Department of Computer Science, California State University Stanislaus, 1 University Circle, Turlock CA 95382 Received 16 April 2021; accepted 20 July 2021 ______________________________________________________________________________ Abstract This paper will cover how I utilized a neural network to train a Jetbot to imitate predator-prey relationships. The Jetbot was created using blueprints provided by NIVDIA. The robot was trained using PyTorch and its various open-source libraries. The goal of this project was to improve upon the research conducted by William Grey Walter and Michael Arbib. Both conducted previous work on replicating animal behavior through predator-prey relationships to examine the interconnectedness of the brain. By replicating their research using a neural network, I can better emulate predator-prey relationships and closely analyze the connectivity of the brain. Keywords: neural network, robotics, biology, machine learning ______________________________________________________________________________ Introduction Robotics has seen a dramatic evolution since This project sought to improve on the its advent. Early robotics saw its place in a child’s previous work done by Walter and Arbib by utilizing playroom as wind-up toys with preprogrammed a neural network. A neural network is a motions (Nocks, 2017). Now robotics can be seen in mathematical-based learning system that uses a hospitals assisting with complex surgeries (Bogue, network of functions to understand various kinds of 2011) or in disaster zones searching for survivors data. The concept originated from human biology and (Tadokoro, 2005). A major contributor to this how the human brain functions using neurons to evolution was the inspiration garnered from interpret inputs from the biological senses biological beings. (Schmidhuber, 2015). Our goal was to mimic Scientists looked to the anatomy of living predator-prey behaviors by training the neural creatures to overcome challenges with robotic network to identify predator versus prey and movement. Big Dog was developed by Boston assigning actions to the different identifications. By Dynamics to overcome problems faced by wheeled doing this, we could observe how the brain interprets robots (Raibert et al., 2008). Another example is situations and executes actions. micro aerial vehicles (MAVs). The downside to using Previous Research these was its inability to land anywhere. Scientists There has been extensive research into looked to flying insects to solve this problem and predator-prey relationships with regard to robotics. developed a MAV that can perch and take off from Two of the most fundamental projects are William any surface and consequently land safely (Granule et Grey Walter’s tortoises and Michael Arbib’s Rana al., 2016). Computatrix. Walter’s tortoises were inspired by Scientists also analyzed behaviors to neuroscience and the complex interconnectedness of understand and implement complex thinking. Two examples are William Grey Walter’s tortoises neurons within the brain (Walter, 1953). Arbib (Walter, 1953) and Michael Arbib’s Rana wanted to improve on Walter’s work by making the Computatrix (Arbib 2003). Walter and Arbib studied thinking process more like the biological life form he predator-prey relationships in various animals and based his robots on (Arbib, 2003). Both scientists replicated these behaviors using the technology were interested in exploring predator-prey behaviors available to them at the time. They believed that by as a means for understanding thinking. studying these relationships they would gain insight William Grey Walter is considered a into the association between brain and action (Arbib, founding father of creating authentic artificial life 2003). Both will be discussed further in the (Holland, 1997). He achieved this honor through the background section of this paper as they are the direct creation of a series of tortoises designed to replicate inspiration for this project. complex behaviors by utilizing a rich interconnectedness between a minimal number of parts (Walter, 1953). Walter’s first design was called biological brain. I trained this neural network to the “Machina Speculatrix” which roughly translates identify a blue pool ball as a predator and a yellow to "machine that thinks" (Mataric, 2007). It was a pool ball as a prey. I then analyzed what this simple design with the only resemblance of a tortoise programmed behavior can teach us about biological being its clear plastic shell protecting the inner behavior. The remainder of this paper will discuss the mechanical components. details of how this was accomplished. There was one driving wheel controlled by a There are several different ways to examine steering and driving motor. Attached to this wheel was a single photoelectric cell. This cell took in the the interconnectivity of the brain. For this project, I light as sensory input and was attached to the driving focused on functional connectivity. Functional wheel to ensure it was always pointing in the connectivity is based on statistics and refers to parts direction the robot was facing. The robot was taught of the brain that are dependent on one another to react to light in specific ways. If the light was of regardless of having direct structural links (Sporns, high intensity, it was programmed to flee. If the light 2007). The living brain relies heavily on statistics to was of low intensity, it was programmed to approach. manage the flow of information gathered from the This mimicked remedial predator-prey behavior environment and determine behavioral output (Chen, exhibited in tortoises and other animals (Walter, 2019). Because of this, neural networks are designed 1953). to make decisions based on statistical inferences to Michael Arbib believed that while Walter’s model the brain as closely as possible. tortoises were a pivotal invention in robotics For the design of the robot, I decided to use evolution, they were not truly “thinking” machines. NVIDIA’s Jetbot. NVIDIA provides all design files He sought to improve on Walter’s research by and a complete list of parts to create the Jetbot from creating his own set of robots exhibiting frog and scratch on GitHub. My mentor, Dr. Xuejun Liang, toad-like visuomotor coordination. He began by followed this guide to construct the Jetbot I used for making robots with basic predator-prey functions, this project. The Jetbot utilizes the computational like those of Walter’s tortoises. A small object would power of the Jetson Nano. This computer allows represent a prey, while a large object would represent users to run multiple neural networks in parallel a predator. If a small moving object were presented, (NVIDIA). I also decided to use the convolutional the robot would snap at the object. If a large moving neural network (CNN) AlexNet. object is in view, the robot would avoid the object. AlexNet is different from earlier CNNs due Arbib took this a step further by relating the behavior to its hierarchical image classification structure. At to the anatomy of a frog (Arbib, 2003). the lower layers of the network, images are seen as Each eye of a frog sends visual information highly pixelated with low resolution. This layer is to the opposite side of the brain, and the midbrain used to determine rough features like the edges of an region known as the tectum. Layered in front of the object or color. At the higher layers, images are tectum is the pretectum. Arbib made the hypothesis processed in greater detail. By performing image that the small moving object identification and the classification in this manner, the computational time resulting action are in the tectum while the large is greatly decreased allowing large datasets to be moving identification and corresponding action are in analyzed and the creation of accurate models to be the pretectum. Through this additional programming, produced (Krizhevsky et. al, 2012). Due to these Arbib observed an interesting behavior. When two advantages, AlexNet was an ideal CNN for the small moving objects were identified, the robot project. would snap in between the two objects rather than I trained the robot to perform three actions, picking one over the other (Arbib, 2003). The same collision avoidance, avoidance of predators, and behavior can be observed in sentient frogs. This attraction to prey. The intention of teaching the robot showed that Arbib was successfully able to mimic collision avoidance was to simulate an animal animal behavior in his Rana Computatrix and create a moving about its territory while looking out for truly thinking machine (Murphy, 2019). predators or prey. I used a table to simulate its Methodology territory and two pool balls to represent predator vs Inspired by these two scientists, I sought to prey. replicate predator-prey behavior to study the I first created categories that separated and interconnectivity of the brain. To improve on their labeled images taken from a mounted camera. These work, I utilized a neural network to better model a categories were: blocked, free, predator, and prey. The categories blocked and free were specific to the Results collision avoidance function. These simply signified The robot was able to successfully execute whether the robot was free to continue its path or if it collision avoidance. It was also able to differentiate should stop to avoid a collision. The categories between predator and prey. When presented with the predator and prey refer to when there is a predator or blue ball, it moved away from it. Thus, showing it prey in the robot’s field of view. recognized it as a predator and avoided it appropriately. When presented with a yellow ball, it To train the model to recognize these moved towards it. Proving that it identified it as a categories, I took two hundred pictures using the prey and responded correctly. mounted camera for each category. These pictures To test the robot’s thinking, different colors were stored in the Jetson nano in their corresponding of pool balls were used.
Details
-
File Typepdf
-
Upload Time-
-
Content LanguagesEnglish
-
Upload UserAnonymous/Not logged-in
-
File Pages4 Page
-
File Size-