Nuclear Environments Inspection with Micro Aerial Vehicles: Algorithms and Experiments

Dinesh Thakur, Giuseppe Loianno, Wenxin Liu, Vijay Kumar

TL;DR

This paper presents a fully autonomous, palm-sized (0.16 m) quadrotor designed to inspect hazardous, confined nuclear environments like damaged Primary Containment Vessels. Running all state estimation, mapping, and obstacle avoidance onboard a computationally limited CPU, the system successfully navigates a full-scale mock-up without GPS or human teleoperation.

What Problem Does This Paper Address?

Decommissioning damaged nuclear reactors (such as those at Fukushima Daiichi) requires surveying highly radioactive, cluttered, and confined spaces. Existing robotic solutions rely on large, heavy ground or aerial vehicles (ranging from 1 to 10 meters and up to 1000 kg) that are remotely teleoperated. These larger robots frequently get stuck in tight areas and rely on continuous communication links.

Furthermore, the interior of a damaged nuclear site lacks GPS, has poor lighting, and features tiny entry hatches (often just 0.1–0.3 m in diameter). The authors aim to bridge this gap by developing a fully autonomous Micro Aerial Vehicle (MAV) small enough to fit through these tiny hatches and capable of navigating and mapping the environment entirely onboard, without relying on a human pilot or external computers.

Methods

The authors built a custom 236 g, 0.16 m diameter quadrotor powered by a Qualcomm Snapdragon Flight board (featuring an ARM processor and a DSP). The onboard software stack is divided into three main components:

  • State Estimation & Control: The drone uses Visual Inertial Odometry (VIO), fusing IMU data with a downward-facing VGA camera (170° field of view). An Extended Kalman Filter (EKF) and Unscented Kalman Filter (UKF) estimate the vehicle's state at 500 Hz, feeding into a nonlinear proportional-derivative controller.
  • Mapping: A front-facing stereo camera generates a dense 3D voxel grid occupancy map. To handle the limited ARM CPU, image rectification is split into separate threads, and a block-matching algorithm extracts disparity maps at 15 Hz to identify obstacles.
  • Planning: The drone follows a pre-generated global trajectory but uses an online reactive planner to avoid unexpected obstacles. It uses Quintic B-splines to smoothly and dynamically deviate around obstacles in real-time, ensuring continuous velocity and acceleration.

Datasets

Not discussed in the paper. The research focuses on real-world physical experiments in a mock-up facility rather than training or evaluating on static datasets.

Key Insights and Ideas

  • Full Autonomy on a Micro-Scale: It is entirely possible to run a complete autonomy stack (VIO, stereo depth mapping, and B-spline trajectory replanning) concurrently on a highly constrained mobile CPU for a sub-250g drone.
  • Lighting is the Limiting Factor: Because the system relies heavily on visual features for odometry (VIO), ambient illumination is a critical failure point. The system can handle moderate low-light but fails catastrophically in near-darkness, highlighting the strict necessity for onboard active lighting in real nuclear deployments.
  • Stereo Vision is Viable for Micro-Drones: Despite the very small baseline (distance between the two lenses) of the micro-drone's stereo camera, it is still accurate enough to detect and avoid relatively thin obstacles in real-time.

Key Results

The system was tested in a full-scale (5 m diameter) mock-up of a Primary Containment Vessel (PCV) pedestal at the Southwest Research Institute (SwRI).

  • Obstacle Avoidance: The MAV successfully detected and avoided unmapped obstacles (like PVC pipes and aluminum trusses) as small as 0.25 m in diameter while flying at a nominal velocity of 0.5 m/s.
  • Luminance Stress Tests: The authors tested the VIO system under varying light conditions. The drone maintained stable tracking at 41.8 lux and 9.5 lux. However, at 0.5 lux, visual features degraded too much, causing VIO drift and a subsequent crash.
  • End-to-End Inspection: In the final test, the drone successfully started outside the mock-up, flew into the cluttered pedestal area, mapped the environment, avoided suspended obstacles, and safely maneuvered out.

Limitations and Future Work

  • Active Lighting: The current platform lacks an onboard light source. Future iterations need an integrated LED payload to operate in environments darker than ~8 lux.
  • Local Minima in Planning: The current reactive planner can get stuck in local minima; integrating a real-time global planner is necessary for navigating more complex, maze-like obstacle fields.
  • Speed Constraints: Faster sensors and algorithms are needed to safely increase the flight speed beyond 0.5 m/s, which is important for time-sensitive, battery-constrained missions.
  • Further Miniaturization: The authors aim to shrink the platform down to 0.1 m in diameter to fit into even tighter access hatches.

Summary updated

Abstract

In this work, we address the estimation, planning, control and mapping problems to allow a small quadrotor to autonomously inspect the interior of hazardous damaged nuclear sites. These algorithms run onboard on a computationally limited CPU. We investigate the effect of varying illumination on the system performance. To the best of our knowledge, this is the first fully autonomous system of this size and scale applied to inspect the interior of a full scale mock-up of a Primary Containment Vessel (PCV). The proposed solution opens up new ways to inspect nuclear reactors and to support nuclear decommissioning, which is well known to be a dangerous, long and tedious process. Experimental results with varying illumination conditions show the ability to navigate a full scale mock-up PCV pedestal and create a map of the environment, while concurrently avoiding obstacles.
BibTeX
@inproceedings{thakur2018nuclear,
  author = {Dinesh Thakur and Giuseppe Loianno and Wenxin Liu and Vijay Kumar},
  title = {{Nuclear Environments Inspection with Micro Aerial Vehicles: Algorithms and Experiments}},
  booktitle = {ISER},
  year = {2018},
  url = {https://openreview.net/forum?id=LZ4NYoMTgU},
}

Loading related content...

Figure