Lighter-Than-Air Autonomous Ball Capture and Scoring Robot - Design, Development, and Deployment

Joseph Prince Mathew, Dinesh Karri, James Yang, Kevin Zhu, Yojan Gautam, Kentaro Nojima-Schmunk, Daigo Shishika, Ningshi Yao, Cameron NowzariarXiv 2023 · DOI · arXiv

Abstract

This paper details the design and deployment of an autonomous Lighter-Than-Air (LTA) robot developed for the "Defend The Republic" (DTR) aerial competition. Utilizing a "Minimum Viable System" (MVS) philosophy, the robot integrates edge AI via YOLOv5 on a Google Coral System-on-Module (SOM) and a custom four-motor configuration to achieve 3D navigation, ball capture, and scoring. We discuss the hardware-software co-design, the importance of in-situ training for robust perception, and the operational results from the 2023 competition.

Introduction

Autonomous LTA robotics presents a unique set of challenges compared to ground or rotor-based systems. Because blimps rely on buoyancy, weight is the ultimate constraint—every additional gram of sensor or battery requires more helium volume, which in turn increases drag and reduces agility. The researchers addressed the challenge of building a fully autonomous agent for the DTR competition in less than a year. The robot needed to solve three major problems simultaneously: 3D locomotion in a three-dimensional space for 30-minute intervals, perception of small, neutrally buoyant balls and specific geometric goals, and physical interaction to capture and "shoot" balls through a goal.

Hardware Architecture

The team adopted a "Minimum Viable System" philosophy, stripping away non-essential components like LiDAR to save weight. The final agent weighed 645g and was supported by a 770-liter helium envelope, providing a 20% lift safety factor. The entire processing unit and controller board were miniaturized to just 72g.

Side and front views of the LTA robot showing the envelope, gondola, and motor configuration.

As shown in Figure 3, the robot features a custom 4-motor configuration. Two motors provide differential drive for yaw (turning), a third handles altitude, and a crucial fourth motor at the rear provides forward thrust. This fourth motor serves a dual purpose: it drives the blimp forward and, when reversed, acts as a "blower" to eject the ball into the goal. By adding this fourth motor, the team decoupled yaw control from forward movement, significantly simplifying the control mathematics and allowing for a more robust and predictable flight path.

Perception and Control

Instead of complex sensor fusion, the team relied on a single monocular USB camera. They deployed a YOLOv5 object detection model quantized to run on a Google Coral SOM, allowing for real-time, on-board inference.

The software uses a subsumption-like architecture—a 4-state finite automata. The robot cycles through four modes: Ball Search, Ball Capture, Goal Search, and Goal Score. Higher-level behaviors, such as capturing a ball, automatically override lower-level ones when a target is detected.

State transition logic for the autonomous agent's four operational modes.

As detailed in the control logic (Figure 4), the system uses specific error parameters ($e_{yaw}, e_{vert}$) to guide the PD (Proportional-Derivative) controllers. These controllers allow the blimp to navigate to targets with a 5-10cm error tolerance.

Datasets and In-Situ Training

The team found that generic datasets were insufficient for the specific lighting and background of the competition arena. They created a custom DTR dataset of 8 classes (green/purple balls and orange/yellow square/circle/triangle goals), requiring between 200 and 400 unique images per class.

A key strategy was "in-situ" training—collecting data directly from the competition venue (Eagle Bank Arena). This allowed the model to learn specific environmental "distractors," such as bright stadium lights that might otherwise be mistaken for game balls. This proved vital for reducing false positives caused by stadium lighting "blooming" effects.

Results and Discussion

The robot successfully performed autonomous "search-capture-score" loops during the April 2023 competition. The team proved that a well-trained vision model could handle "holding" detection by analyzing pixel colors near the cage, justifying the removal of weight-heavy LiDAR sensors.

However, several limitations remain. The vision models are highly tuned to specific arena conditions, requiring re-training for new venues. Additionally, the current envelope design lacks control surfaces like fins, making it highly dependent on motor thrust for stability. Future work will focus on self-tuning methods to account for physical idiosyncrasies in hand-assembled units and swarm intelligence to prevent multiple agents from competing for the same ball.

Summary updated

Abstract

This paper describes the full end-to-end design of our primary scoring agent in an aerial autonomous robotics competition from April 2023. As open-ended robotics competitions become more popular, we wish to begin documenting successful team designs and approaches. The intended audience of this paper is not only any future or potential participant in this particular national Defend The Republic (DTR) competition, but rather anyone thinking about designing their first robot or system to be entered in a competition with clear goals. Future DTR participants can and should either build on the ideas here, or find new alternate strategies that can defeat the most successful design last time. For non-DTR participants but students interested in robotics competitions, identifying the minimum viable system needed to be competitive is still important in helping manage time and prioritizing tasks that are crucial to competition success first.
BibTeX
@misc{mathew2023lighter,
  author = {Joseph Prince Mathew and Dinesh Karri and James Yang and Kevin Zhu and Yojan Gautam and Kentaro Nojima-Schmunk and Daigo Shishika and Ningshi Yao and Cameron Nowzari},
  title = {{Lighter-Than-Air Autonomous Ball Capture and Scoring Robot - Design, Development, and Deployment}},
  year = {2023},
  eprint = {2309.06352},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2309.06352},
}
Figure