Abstract
This paper details a 14-week multidisciplinary internship program designed to move beyond "toy" projects and create a functional Research and Development (R&D) ecosystem for autonomous systems. By pairing undergraduate interns with professional mentors, the program successfully developed a modular pipeline that bridges the gap between bio-plausible simulation and physical robotic hardware. The study evaluates the feasibility of novice teams implementing complex AI concepts—such as Spiking Neural Networks (SNNs) and Bayesian reasoning—on physical platforms within a single summer.
Introduction
The field of autonomous systems is increasingly interdisciplinary, requiring expertise in AI, biology, psychology, and robotics. However, traditional engineering education often relies on "chalk and talk" lectures or highly structured lab exercises with predetermined outcomes. The authors identify a gap in workforce development: the lack of a research-based internship model that balances education with actual R&D progress. This program sought to determine if a novice, multidisciplinary team could quickly learn complex concepts and implement them on a physical robot.
Methodology
The internship followed a modular, three-phase framework over 14 weeks:
- Training (4 weeks): Interns completed tutorials on the Robot Operating System (ROS), Modeling and Simulation (M&S) tools (V-REP, Gazebo), and the Nengo neural modeling framework.
- Innovation (4 weeks): The team developed a "reasoner" module. They explored two parallel approaches: a Bayesian Belief Network (using ProbLog) and a bio-plausible Basal Ganglia model (using Nengo and Spiking Neural Networks).
- Integration (5 weeks): The "reasoner" was integrated into a feedback loop (Observe-Orient-Decide-Act) and deployed on a TurtleBot2 outfitted with a Jetson TX2 and various cameras.
The project utilized Agile Software Development principles, allowing for iterative improvements and "fail-fast" learning. This modularity allowed the team to swap reasoning components without rebuilding the entire perception or navigation stack.
Implementation and Key Insights
The "perceptor" module utilized a TensorFlow-based Convolutional Neural Network (CNN) trained on the COCO dataset, enabling the robot to identify objects like cups, books, and people. A key insight from the program was the "Power of Multidisciplinarity." Including a Biology major helped structure the neural models to incorporate "amygdala-like" functionality. This allowed the robot to have "Moods" (Hunger, Fear, Curiosity) that influenced its "Intentions" (Feed, Flee, Explore).
However, the team also encountered the "Integration Tax." The authors found that "data wrangling" and software integration (connecting Python/Nengo to ROS) occupied the majority of the team's time. This highlights that in real-world robotics, the "plumbing" is often as challenging as the "brain."
Results
The team successfully created an end-to-end M&S-to-hardware pipeline. The robot demonstrated the ability to change its behavior based on simulated "moods." For example, high confidence in a "cup" detection might trigger a "Hunger" mood, leading to a "Feed" intention (moving toward the object).
Furthermore, the project proved that Spiking Neural Networks (3rd generation AI) could be effectively run on Commercial Off-The-Shelf (COTS) hardware like the NVIDIA Jetson TX2 in a robotics context. Five undergraduates with no prior autonomy experience were able to contribute to AFRL-level research questions within 14 weeks.
Discussion and Future Work
The authors noted several "lessons learned" for future iterations. Mentors initially withheld the "end goal" to avoid overwhelming interns, but the interns later reported they would have preferred knowing the full scope earlier to better plan their work. Future programs will involve more rigorous vetting of "ancillary tools" (like the perceptors) before the interns arrive to prevent wasting time on low-level software bugs. Additionally, there is a need for better quantitative metrics to evaluate autonomous performance beyond simple "proof of concept" demonstrations.
Summary updated