AgiPIX: Bridging Simulation and Reality in Indoor Aerial Inspection

1. Problem Statement

Autonomous aerial inspection of critical assets (e.g., nuclear power plants, warehouses, tunnels) faces significant challenges in indoor, GPS-denied environments. These include:

• Operational Constraints: Limited space, poor lighting, and the need for reliable close-proximity operation around obstacles.
• Technological Gaps: Existing commercial drones are often too large, proprietary, or lack the onboard compute for advanced learning-based perception. Conversely, open-source research platforms often lack compactness, high thrust-to-weight ratios (TWR), or synchronized active sensing (LiDAR/IMU).
• Sim-to-Real Transfer: There is a lack of a unified, open-source platform that allows for rapid iteration by transferring autonomy components seamlessly between photorealistic simulation and real-world flight without retraining or major code changes.

2. Methodology: The AgiPIX Platform

The authors propose AgiPIX, a co-designed open hardware and software ecosystem comprising four integrated subsystems:

A. AgiREAL (Hardware Platform)

• Design: A compact, carbon-fiber quadrotor with a diagonal span of ~495 mm (including guards), designed to pass through standard 800 mm doors.
• Performance: Achieves a static thrust-to-weight ratio of 3.5:1, enabling agile flight and close-proximity operation.
• Sensing & Synchronization:
  • Primary Sensors: Livox MID-360 3D LiDAR and a triple-redundant IMU (Pixhawk Orange).
  • Synchronization: Uses a Pulse-Per-Second (PPS) signal generated by an ESP32 (synchronized to the Pixhawk master clock) to ensure hardware-level time alignment between the LiDAR and IMU, critical for accurate LiDAR-Inertial Odometry (LIO).
  • Compute: Equipped with an NVIDIA Jetson Orin NX (Super) (16 GB RAM, 157 TOPS) for onboard GPU-accelerated processing.
  • Modularity: Supports optional sensors like RealSense depth cameras, RGB cameras, and radiation detectors.

B. AgiAUTO (Modular Software Stack)

• Architecture: A containerized ROS 2 autonomy stack running on the companion computer, interfacing with the low-level PX4 flight controller via Micro XRCE-DDS.
• Key Modules:
  • Perception: Uses a modified Adaptive-LIO for state estimation and mapping in GPS-denied environments. It employs LV-DOT for dynamic obstacle flagging.
  • Planning & Control: Utilizes a local ViGO planner for trajectory generation and an on-manifold MPC for tracking. It supports multiple flight modes: Twist control, Goal control, Navigation, and Exploration (closest-frontier).
  • Learning: Designed to support data-driven policies. The paper demonstrates the deployment of FAINT (Fast Appearance-Invariant Navigation Transformer), a learning-based policy trained in simulation and transferred zero-shot to the real drone.
  • Logging: Uses agi_logger with .mcap format for reliable, segmented, and lossless data recording.

C. AgiSIM (Digital Twin)

• Environment: A photorealistic simulation built on NVIDIA Isaac Sim.
• Parity: Enforces strict configuration parity between simulation and reality at three levels:
  1. Sensor Parity: Matching calibration, noise models, and topic conventions.
  2. Timing Parity: Consistent propagation of simulation time to ROS 2.
  3. Software Parity: Running identical container images and launch files in both environments.
• Integration: Uses the Pegasus interface to bridge PX4 SITL (Software-in-the-Loop) with Isaac Sim, allowing the same ROS 2 autonomy stack to run in both domains.

D. AgiUI (User Interface)

• Function: A lightweight web-based interface for mission control and monitoring.
• Communication: Uses low-bandwidth MQTT to transmit state, maps, POV imagery, and payload data.
• Features: Includes fleet overview, synchronized 3D mapping views, radiation heatmaps (for nuclear inspection), and Human-Robot Interaction (HRI) monitoring (tracking operator cognitive load via physiological data).

3. Key Contributions

1. AgiREAL: An open-source, compact, high-performance UAV with hardware-synchronized 3D LiDAR and onboard GPU compute, optimized for agile indoor flight.
2. AgiAUTO: A modular, containerized ROS 2 autonomy stack that enables rapid deployment and supports both classical and learning-based navigation.
3. AgiSIM: A photorealistic Isaac Sim digital twin that shares the exact software stack with the real hardware, facilitating zero-shot sim-to-real transfer.
4. AgiUI: An integrated operator interface for mission control, visualization, and HRI monitoring.
5. Open Release: Full release of hardware designs, simulation assets, containerized software, and documentation to lower barriers for research.

4. Experimental Results

The platform was validated through simulation and real-world experiments, including participation in the ENRICH 2025 robotics hackathon at the decommissioned AKW Zwentendorf nuclear power plant.

• Trajectory Tracking:
  • Tested on "Lemniscate" and "Up-Down Spiral" paths at 6 m/s.
  • AgiSIM RMSE: 0.0637 m (Lemniscate) / 0.0368 m (Spiral).
  • AgiREAL RMSE: 0.1404 m (Lemniscate) / 0.1049 m (Spiral).
  • Result: Sub-decimeter accuracy in simulation and low-decimeter accuracy in reality, demonstrating stability despite unmodeled disturbances.
• Mapping & Exploration:
  • Achieved an Intersection over Union (IoU) of 0.96 against ground-truth maps in a representative indoor environment.
  • Successfully mapped 72% of a planned flight distance in a large-scale nuclear plant environment before a communication failure.
• Learning-Based Navigation:
  • Deployed the FAINT policy (trained in sim) on the real drone without retraining.
  • Achieved 1.07 m RMSE on a successful run and 1.42 m RMSE on a challenging route, with 120 ms inference latency on the Jetson Orin NX.
• System Utilization:
  • The system utilized only 41% of CPU and 37% of GPU resources, leaving significant headroom (59% CPU, 63% GPU) for future scalability and complex algorithms.

5. Significance

AgiPIX addresses a critical bottleneck in aerial robotics research: the disconnect between simulation and reality. By providing a co-designed, open, and reproducible platform, it enables:

• Rapid Iteration: Researchers can test new algorithms in a photorealistic digital twin and deploy them immediately on real hardware with minimal tuning.
• Accessibility: It lowers the financial and engineering barriers for developing advanced indoor inspection systems, particularly for hazardous environments like nuclear facilities.
• Validation: It serves as a robust benchmark for evaluating state-of-the-art perception, planning, and learning-based navigation methods in realistic, GPS-denied scenarios.

The platform successfully demonstrated its utility in a high-stakes, real-world environment (nuclear power plant inspection), proving that open-source, containerized autonomy stacks can achieve performance comparable to proprietary systems while offering superior flexibility and reproducibility.