The History of the RIT Pacbot Team

The Rochester Institute of Technology (RIT) Pacbot team is a part of the RIT Robotics Club, also known as MDRC. The team participates in the annual Pacbot competition, where autonomous robots navigate a maze to collect pellets and evade or capture ghosts. This document chronicles the history of the RIT Pacbot team, drawing from a team meeting held on May 2, 2025, following the competition hosted at RIT on April 19, 2025.

The RIT Pacbot team's participation in the competition dates back to its early years. Quinn Tucker, who joined RIT in the fall of 2019, recalled being told about the team's unexpected victory in the 2019 Pacbot competition. Ben Giacalone, also a freshman in 2019, corroborated this, describing it as a win achieved through "complete accent." At that time, the robots from various teams were generally unreliable, and their artificial intelligence (AI) capabilities were rudimentary, primarily relying on heuristic approaches. The RIT robot that secured the first-place finish was described as being hastily assembled the night before the competition and barely functional.

Following this surprising success, the team faced challenges in subsequent years. The onset of the COVID-19 pandemic in 2020 significantly disrupted the competition schedule, with the Pacbot competition being canceled in 2020 and likely in 2021 as well. This period of uncertainty led to demotivation among some team members. Despite the lack of competitions, the team continued to work on their robot and AI, though progress on the physical robot was slow, with early prototypes being simple box-shaped designs that lacked effective movement.

The team's efforts gained renewed momentum in 2023 when the Pacbot competition was reinstated. Ben Giacalone, who had been exploring reinforcement learning (RL) algorithms, rejoined the team with the goal of applying these techniques to the robot's high-level AI. After considerable experimentation, the team achieved a breakthrough, developing an RL-based AI that outperformed their previous heuristic methods. This marked a significant advancement in the team's approach to the competition.

In the fall of 2023, Quinn Tucker undertook an independent study under Dr. Zhao, focusing on implementing various RL algorithms for the Pacbot challenge. This effort, partly motivated by gaining academic credit for his work on the team, saw the exploration of Deep Q Network (DQN), Proximal Policy Optimization (PPO), and AlphaZero algorithms. While AlphaZero proved computationally intensive, the DQN agent ultimately demonstrated the best performance and became the foundation for the team's high-level policy in subsequent competitions. Ben Giacalone played a crucial role in guiding Quinn with the fundamentals of reinforcement learning during this period.

Michael Elia joined the team as a freshman in 2022. Under the leadership of Christian Guerrero, the team initially focused on heuristic-based AI. However, progress in this area was disappointing, with the models frequently encountering simple scenarios they couldn't navigate effectively. The team did attempt to build a robot that year, but it remained unfinished due to incomplete electronics.

The 2023-2024 academic year saw Michael, Quinn, and Ben Giacalone, all software-focused, attempt to build a robot. This endeavor proved unsuccessful, with the resulting robot unable to move. This experience highlighted the necessity of dedicated hardware expertise within the team.

The arrival of Brian Doolin, with a background in BattleBots, proved to be a turning point for the team's hardware capabilities. Brian introduced Onshape, a collaborative CAD software, to streamline the robot design process. He was tasked with creating the smallest possible robot, leading to the current three-wheeled omni-drive design, which fits within a five-inch diameter circle. This design prioritized compactness to navigate the seven-inch wide passages of the competition arena, a lesson learned from past robots getting stuck. The robot features a top shell, a mount for a screen (used for software debugging), a battery, and three motors in a triangular configuration to ensure consistent ground contact and encoder accuracy. The team utilizes Pico boards and a custom-designed PCB, referred to as a "hat," for motor control, power distribution, and sensor connections. This PCB, designed by Michael Elia using EasyEDA software, significantly improved the organization and reduced the size of the robot's electronics compared to earlier wire-based setups and proto boards.

The robot is equipped with several types of sensors. Encoders on each wheel provide data on wheel speed and position. Initially, the team used infrared distance sensors, which were unreliable. They then transitioned to time-of-flight sensors, which offered improved accuracy. The current model of time-of-flight sensors can detect objects as close as one millimeter, though their effective range is limited by the cone of light they emit hitting the arena floor. To address the robot's orientation, the team incorporated an Inertial Measurement Unit (IMU) featuring an accelerometer, gyroscope, and magnetometer. While the magnetometer proved too noisy for reliable use, the accelerometer and gyroscope provide accurate rotational data, crucial for localization.

Localization, determining the robot's position within the maze, has been a significant challenge. In 2023-2024, Quinn Tucker developed a particle filter-based localization system implemented in Rust. While theoretically sound, this approach proved computationally expensive and struggled with inaccuracies stemming from the computer vision (CV) system's lag in tracking the robot's position. Despite this limitation, the robot performed well in simulation and even achieved second place in one practice run at the 2024 competition, scoring higher than any previous team. However, the CV lag hampered its performance in official runs.

Following the 2024 competition, Michael Elia initiated a significant restructuring of the team's codebase, breaking it into modular crates within the MDRC Pacbot repository on GitHub. The particle filter was not ported to this new system due to its performance issues. Instead, the team shifted towards a stateless localization approach for the 2024-2025 season. Ben Brodeur played a key role in developing this new system, which initially involved raycasting based on CV location and sensor readings. However, this was prone to errors when the robot's rotation deviated. The final stateless localization method adopted involved comparing sensor readings to expected values at discrete locations within the maze grid and selecting the most likely location based on proximity to the CV data and a point system. This simpler, computationally efficient algorithm could run directly on the robot, eliminating the delay associated with off-robot processing.

The 2025 competition, hosted at RIT, saw the team achieve first place, with UIUC, who won their own competition, not attending the RIT event. The team also showcased their custom-built competition field, constructed from MDF boards using the club's CNC router.

Looking to the future, Ben Brodeur, the team lead for the next year, outlined several areas for improvement. Enhancing the accuracy of localization remains a priority, potentially by re-exploring state-based methods to address sensor range limitations in long corridors. Improving low-level motor control by incorporating feed-forward terms alongside PID control is also seen as important for smoother movements and preventing the robot from getting stuck.

The team also discussed potential hardware upgrades. While a suspension system was deemed overly complex, exploring flexible 3D-printed mounts or rubber dampeners could help mitigate vibrations. Considering motors with perpendicular axles could further reduce the robot's footprint. The reliability of the current brushed DC motors, which occasionally fail due to brush disintegration, is also a factor to consider for future iterations.

A recurring "bad idea," jokingly raised by Quinn Tucker, was to train the reinforcement learning model directly on the physical robot instead of in simulation, a proposition deemed impractical due to the potential for damage and inefficiency. Another explored but ultimately discarded concept was to have the reinforcement learning model control the low-level motor commands directly. The consensus was that hand-coded motor control, possibly incorporating model-based control techniques, would be more effective.

The team acknowledges the significant progress made over the years, from a barely functional robot to a competition-winning design and the ability to host their own event. The current success is attributed to the dedication and hard work of numerous team members, particularly the graduating seniors Michael Elia and Brian Doolin, as well as the contributions of alumni like Quinn Tucker and Ben Giacalone. As Michael Elia prepares to hand over the team leadership to Ben Brodeur, there is a sense of pride in past achievements and optimism for the future of the RIT Pacbot team. The wealth of knowledge and experience accumulated within the team, documented in their GitHub repository and the memories of its members, will serve as a valuable resource for future generations of RIT students who take on the challenge of autonomous Pacbot navigation.