Team NimbRo (University of Bonn) demonstrated a highly successful approach to autonomous field robotics at the 2020 Mohamed Bin Zayed International Robotics Challenge (MBZIRC). By deploying a fleet of four specialized Unmanned Aerial Vehicles (UAVs) supported by a unified, modular software stack, the team secured second place in the Grand Challenge. This paper summarizes their hardware innovations, including passive mechanical compliance, and their software strategies, such as universal model predictive control and synthetic data training for perception.
Introduction
The MBZIRC competition presents a rigorous testbed for autonomous systems, requiring robots to operate in uncontrolled outdoor environments characterized by harsh lighting, wind, and unreliable GPS. Team NimbRo addressed four primary tasks: high-speed balloon popping, intercepting a moving target, autonomous wall building, and fire fighting. These tasks demand a combination of agility, precise perception, and robust aerial manipulation.
Methodology and Hardware
The team's success was rooted in a modular software stack built on the Robot Operating System (ROS). Key components included a model-independent Model Predictive Controller (MPC), which allowed for consistent flight performance across different drone configurations and changing payloads without manual retuning. To ensure logical consistency, the nimbro_fsm2 state-machine library utilized C++ metaprogramming to verify robot logic at compile-time.
To address the specific requirements of the competition, the team developed four distinct UAVs:
- Jelly: Designed for balloon hunting with detachable "spiked tentacles."
- Chaser: Optimized for target pursuit with a spring-loaded foldable net.
- Lofty: Built for wall construction using a magnetic gripper with passive mechanical compliance.
- Splasher: Equipped for fire fighting with integrated LiDAR, thermal cameras, and a water nozzle.
Robust Autonomy and Perception
Maintaining operational reliability in the field required innovations in both communication and perception. The team developed the nimbro_network package, which employed Forward Error Correction (FEC) to maintain stable data streams over unreliable WiFi, essential for remote monitoring and state estimation.
For perception, the team leveraged synthetic data to train deep learning models. Using the easy_pbr physically-based renderer, they generated approximately 10,000–20,000 synthetic images of targets under various lighting conditions. These models were then fine-tuned with a smaller set of real-world data collected on-site. High-speed onboard inference (30Hz) was enabled by Google Edge TPU accelerators, allowing for reactive flight.
Key Insights: Mechanical Intelligence and Relative Navigation
A major takeaway from the competition was the value of "mechanical intelligence." In the wall-building task, rather than striving for perfect flight stability in windy conditions, the team designed a gripper with passive mechanical compliance (telescopic rods). This allowed bricks to self-align during placement, compensating for slight positioning errors.
Additionally, the team prioritized relative navigation over global GPS coordinates. By steering the robots based on the target's position within the camera or LiDAR frame, they successfully bypassed the significant GPS drift issues typically encountered near large structures.
Results and Lessons Learned
Team NimbRo achieved 2nd place overall in the Grand Challenge. Notable successes included a 92.3% detection accuracy for single balloons and the successful puncture of all balloons during the Grand Challenge run. In fire fighting tests, the system could detect and engage a heat source within 6 seconds of first detection.
However, the competition also highlighted areas for improvement. GPS drift during initialization remained a vulnerability, and "rotor wash" from the propellers occasionally displaced bricks during the wall-building task. Future work will focus on initialization-free localization and more robust reactive obstacle avoidance to handle complex arena geometries.
Reference: Beul, M., et al. (2022). "Target Chase, Wall Building, and Fire Fighting: Autonomous UAVs of Team NimbRo at MBZIRC 2020." Field Robotics.
Summary updated