On Your Own: Pro-level Autonomous Drone Racing in Uninstrumented Arenas

Basic Information

• Paper ID: 2510.13644
• Title: On Your Own: Pro-level Autonomous Drone Racing in Uninstrumented Arenas
• Authors: Michael Bosello, Flavio Pinzarrone, Sara Kiade, Davide Aguiari, Yvo Keuter, Aaesha AlShehhi, Gyordan Caminati, Kei Long Wong, Ka Seng Chou, Junaid Halepota, Fares Alneyadi, Jacopo Panerati, Giovanni Pau
• Category: cs.RO (Robotics)
• Publication Date: October 15, 2025
• Paper Link: https://arxiv.org/abs/2510.13644

Abstract

Unmanned aerial vehicle (UAV) technology is rapidly advancing across multiple industries including agriculture, logistics, defense, infrastructure, and environmental monitoring. Vision-based autonomy is a key enabling factor, particularly for real-world applications. This is critical for operating in novel, unstructured environments where traditional navigation methods may be unavailable. Autonomous drone racing has emerged as a de facto benchmark for such systems. Recent research demonstrates that autonomous systems can exceed human-level performance in racing scenarios. However, direct application to commercial and field operations remains limited because current systems are typically trained and evaluated in highly controlled environments. This paper analyzes and demonstrates system capabilities in both controlled environments (with external tracking available for ground truth comparison) and challenging uninstrumented environments (where no ground truth measurements are available). The study shows that the approach can match professional human pilot performance in both scenarios.

Research Background and Motivation

1. Problem to be Addressed: While existing autonomous drone racing systems can exceed human performance in controlled environments, they face challenges in practical applications, particularly in uninstrumented environments lacking external tracking systems.
2. Problem Significance:
   • Widespread UAV applications across industries require reliable autonomy in unstructured environments
   • Real-world deployments typically lack precise external positioning systems
   • Need to verify robustness of autonomous systems under actual conditions
3. Limitations of Existing Approaches:
   • Dependence on highly controlled environments and external tracking systems
   • Requirement for ground truth data for system fine-tuning
   • Instability in varying lighting conditions and unknown environments
4. Research Motivation: Develop autonomous drone systems capable of achieving professional-level performance in uninstrumented environments, advancing technology toward practical commercial applications.

Core Contributions

1. Achieved Professional-level Autonomous Drone Racing: Attained professional-level performance in both controlled environments (with external tracking) and uninstrumented environments (without ground truth measurements)
2. Proposed Robust Perception and Control Stack: Does not require ground truth data for residual estimation fine-tuning and demonstrates adaptability to multiple lighting conditions
3. Released Professional-grade Flight Dataset: Contains 6 flights from world champion pilots, totaling 240.77 seconds of flight time, 2342.98 meters of flight distance, and maximum speed of 21.29 m/s
4. Verified Human-Machine Competitive Performance: Direct competition with world-class pilots across multiple scenarios, demonstrating system practicality

Methodology Details

Task Definition

Input: Stereo camera image stream, IMU data, race gate position information Output: Drone control commands (collective thrust and body angular rates) Constraints: Real-time requirements, dynamic limitations, obstacle avoidance requirements

Model Architecture

1. Vision Stack

• Gate Detection: YOLOv8n model (3.2M parameters) for race gate detection
• Corner Detection: Improved MobileNetV3-Small model (1.1M parameters) for detecting four inner corners of gates
• Optimization Strategy:
  • Conversion to ONNX graphs and TensorRT engines
  • FP16 precision acceleration
  • Per-frame latency of 24-30ms

2. State Estimation Stack

• VIO Foundation: Intel T265 stereo camera provides visual-inertial odometry
• Drift Correction:

  State vector: x = p_d^T ∈ R³ (position drift vector)
  State propagation: x_{k+1} = Fx_k, P_{k+1} = FP_kF^T + Q
  Kalman update: K_k = P_k^-H^T(HP_k^-H^T + R)^{-1}
  

• IMU Fusion: Extended Kalman filter fusing 500Hz IMU data

3. Control Stack

• Time-Optimal Trajectory Generation: Considering rigid body dynamics and actuator constraints
• Model Predictive Control: Based on PAMPC framework with perception-aware objective disabled
• Latency Compensation: Integrated state predictor compensating for computational and execution delays

Technical Innovations

1. Ground Truth-Free Fine-tuning: Unlike existing methods, the system does not depend on external tracking data for state estimation fine-tuning
2. High-Frequency IMU Integration: Optimized MSP protocol achieving 500Hz IMU data reading, significantly improving upon 10Hz SBUS protocol
3. Robust Vision Processing:
   • Fixed exposure settings reducing motion blur
   • Model distillation approach reducing annotation requirements (only 80 frames manual annotation needed)
4. Real-time Performance Optimization:
   • Real-time Linux kernel configuration
   • GPU-accelerated inference
   • Optimized data flow architecture

Experimental Setup

Datasets

1. Instrumented Track:
   • Reconstructed from RATM dataset
   • 32-camera Qualisys MoCap system providing ground truth
   • Includes sharp turns, spiral segments, and Split-S maneuvers
2. Uninstrumented Track:
   • Reconstructed Track Split-S course
   • Total station positioning (centimeter-level accuracy)
   • Natural lighting variation conditions

Evaluation Metrics

• Lap Time: Time to complete a single lap
• Maximum Speed: Peak velocity achieved during flight
• Path Length: Actual flight trajectory length
• Consistency: Standard deviation across multiple flights
• Reliability: Success rate and collision count

Comparison Methods

• Professional Pilots: 3 professional pilots, including world champion MCK
• External Tracking: Autonomous flight using MoCap system
• Onboard-Only: Autonomous flight using only onboard sensors

Implementation Details

• Hardware Platform: NVIDIA Orin NX + Intel RealSense T265
• Thrust-to-Weight Ratio: ~7:1 (full battery capacity)
• Weight: 665.5g (excluding battery)
• Communication: 1MBaud MSP serial connection

Experimental Results

Main Results

Instrumented Track Performance

Uninstrumented Track Performance

Ablation Studies

1. VIO vs MoCap: Using onboard VIO only, average lap time is merely 4.7% slower compared to external tracking
2. Drift Correction Effect: Kalman filtering significantly improves position estimation accuracy during extended flights
3. IMU Fusion Contribution: 500Hz IMU data fusion provides smoother state estimation

Case Analysis

• Split-S Maneuver: Autonomous system excels in constrained spaces with superior trajectory consistency compared to human pilots
• Spiral Segment: Identified by human pilots as critical performance area; autonomous system achieves competitive performance through trajectory optimization
• Hairpin Turn: Becomes primary limiting factor for autonomous system, requiring conservative thrust-to-weight ratio settings

Experimental Findings

1. Consistency Advantage: Autonomous system demonstrates significantly better consistency (smaller standard deviation)
2. Environmental Adaptability: System successfully adapts to different lighting conditions and track layouts
3. Human-Machine Interaction Challenges: In shared track competition, autonomous system is more vulnerable to collisions

Related Work

Main Research Directions

1. AlphaPilot Challenge (2019): Pioneering AI drone racing competition
2. Deep Reinforcement Learning Methods: Kaufmann et al. demonstrated superhuman performance in 2023
3. Dataset Construction: RATM dataset provides benchmarks for algorithm development

Advantages of This Work

• Real Environment Validation: First to achieve professional-level performance in uninstrumented environments
• Practical Orientation: Does not depend on external tracking systems, closer to real-world applications
• System Completeness: Provides complete solution from perception to control

Conclusions and Discussion

Main Conclusions

1. Autonomous drone systems can achieve professional pilot-level performance in uninstrumented environments
2. Appropriate engineering optimization and system integration are more important than complex algorithms
3. Consistency is the primary advantage of autonomous systems relative to humans

Limitations

1. Shared Space Challenges: Insufficient adaptability in human-machine hybrid racing
2. Environmental Generalization: Still requires limited data for environmental adaptation
3. Peak Performance: Slightly inferior to top pilots in best single-lap times

Future Directions

1. Transition from stereo cameras to monocular cameras, more closely mimicking human vision
2. Improve multi-agent interaction and collision avoidance
3. Enhance sim-to-real transfer capabilities

In-Depth Evaluation

Strengths

1. High Practical Value: Addresses critical gap from laboratory to real-world applications
2. Engineering Completeness: Provides detailed hardware and software implementation details
3. Comprehensive Evaluation: Includes multi-dimensional quantitative and qualitative assessment
4. Data Openness: Publicly releases high-quality flight dataset

Weaknesses

1. Limited Algorithmic Innovation: Primarily engineering integration of existing techniques
2. Insufficient Theoretical Analysis: Lacks theoretical analysis of system performance boundaries
3. Scenario Limitations: Validation only in indoor structured racing tracks

Impact

1. Promotes Industrialization: Provides important reference for commercialization of autonomous UAV technology
2. Benchmark Significance: Establishes performance baseline in uninstrumented environments
3. Open-Source Contribution: Dataset and code release will advance field development

Applicable Scenarios

• Indoor warehouse and logistics applications
• Infrastructure inspection
• Search and rescue missions
• Entertainment and sports competition

References

1 Hanover, D., et al. "Autonomous drone racing: A survey." IEEE Transactions on Robotics, 2024. 2 Kaufmann, E., et al. "Champion-level drone racing using deep reinforcement learning." Nature, 2023. 3 Bosello, M., et al. "Race against the machine: A fully-annotated, open-design dataset." IEEE RAL, 2024.

Overall Assessment: This is a practically valuable engineering-oriented paper that successfully translates laboratory technology into a deployable real-world system. While relatively limited in algorithmic innovation, its contributions in real-world validation and systems engineering are significant for advancing the industrialization of autonomous UAV technology.