⚡ electrical engineering

RADAR Perception for Dynamic Obstacle Avoidance onboard small-scale Quadrotor UAVs

This paper presents the first onboard mmWave RADAR-based perception and control system for small-scale quadrotor UAVs that achieves fast dynamic obstacle avoidance with low-latency (14 ms) performance, validated by 390 experiments demonstrating high accuracy across varying object sizes, lighting conditions, and smoke.

Original authors: Dnyandeep Mandaokar, Bernhard Rinner

Published 2026-08-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Dnyandeep Mandaokar, Bernhard Rinner

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a world where tiny flying robots, like mechanical hummingbirds, zip through crowded rooms to deliver medicine or explore disaster zones. For these drones to be truly useful, they can't just follow a pre-programmed path; they need to be agile enough to dodge a falling branch, a swerving person, or a sudden gust of wind without crashing. This is the challenge of "Dynamic Obstacle Avoidance." Currently, most drones rely on cameras to see the world, much like humans do. But cameras have a fatal flaw: they go blind in the dark, get confused by thick smoke, and struggle to tell you how fast something is moving without doing heavy math. Enter the radar. You might know radar from air traffic control or weather stations, but it's essentially a super-sonic flashlight that bounces radio waves off objects to tell you exactly where they are and how fast they are zooming, regardless of whether it's pitch black or filled with fog. The big question scientists have been asking is: Can we shrink this powerful, fog-proof radar down to fit on a tiny, battery-powered drone and make it fast enough to save the drone's life in a split second?

This paper says "yes," but with a few important caveats about how small the objects can be. The researchers built a system that lets a small drone use a millimeter-wave radar to see obstacles, track them, and swerve out of the way—all while flying. They didn't just simulate this on a computer; they strapped the radar to a real drone, put the brain of the operation on a tiny computer (a Raspberry Pi 4B) that fits in a backpack, and threw balls at it in various conditions. The result is a "perception-and-control" system that acts like a reflex arc for the drone. Instead of taking a long time to think, "Oh, a ball is coming, let me calculate a new path," the system reacts almost instantly, calculating a safe acceleration to dodge the threat.

The team found that their radar-based system works remarkably well in challenging environments, though with some limits on object size. In experiments involving 390 throws of balls of different sizes, the system successfully detected and avoided obstacles in bright light and total darkness. In thick smoke, the system maintained high performance for larger balls, but the avoidance rate dropped to 70% for the smallest tennis balls, highlighting that very small, non-metallic objects can still be tricky to spot in poor visibility. The drone's "brain" was able to process the radar data and issue a command to dodge in about 14 milliseconds. To put that in perspective, a human blink takes about 300 milliseconds; this drone reacts more than 20 times faster than you can blink. In terms of accuracy, the drone kept its position within about 0.15 meters in the horizontal direction and less than a meter in the vertical direction, which is plenty of room to avoid a collision.

However, the paper is very clear about what this system cannot do yet. It explicitly rules out the idea that this radar setup is perfect for every single scenario. The system struggles with very small, non-metallic objects (like a tiny tennis ball) when they are far away or in smoke, because the radar waves don't bounce off them strongly enough to be seen. If the drone doesn't see the object until it's too close, the math says a crash is inevitable. The researchers also note that while the system is fast, it relies on the computer not getting "sped up" by too many other tasks; if the computer gets bogged down, the reaction time slows, and the safety margin shrinks.

The core discovery is that millimeter-wave radar is a viable, robust alternative to cameras for fast drone dodging, provided the obstacles are large enough to be seen. The system uses a clever "safety bubble" concept. Imagine the drone and the obstacle are surrounded by invisible, expanding circles. The system calculates a "critical distance"—a point of no return. If the obstacle crosses this line, the drone must have already started dodging. By measuring how fast the obstacle is moving and how fast the drone can turn, the system knows exactly how early it needs to trigger the dodge. They proved that with their setup, the drone could trigger a dodge when the obstacle was still roughly 2 to 4 meters away, giving it plenty of time to swerve safely.

In the end, this research suggests that we don't need expensive, heavy, or power-hungry computers to make drones safe in the dark or smoky. A lightweight radar and a simple algorithm can do the heavy lifting. The paper concludes that while there is still work to be done to handle smaller objects and faster speeds, this approach offers a strong, practical trade-off between speed, safety, and the limited power available on small flying robots. It's a step toward a future where drones can fly through a storm of debris or a dark warehouse and still find their way without crashing.

Technical Summary: RADAR Perception for Dynamic Obstacle Avoidance onboard Small-Scale Quadrotor UAVs

Problem Statement
Fast dynamic obstacle avoidance (DOA) for unmanned aerial vehicles (UAVs) requires a perception system that balances low latency, sufficient sensing range, and robustness in adverse environmental conditions. Current state-of-the-art DOA systems rely heavily on visual sensors (cameras) or LiDAR. However, optical sensors struggle with poor lighting, smoke, and reflective surfaces, while also lacking direct velocity measurements, which necessitates additional processing and increases latency. LiDAR offers long-range detection but demands high computational resources and also fails to directly measure velocity. Millimeter-wave (mmWave) RADAR offers a cost-effective alternative with direct range, velocity, and position measurements, and robustness to lighting and smoke. However, RADAR data is sparse, making object size estimation difficult, and its application for fast onboard DOA (exceeding 50 Hz) on resource-constrained platforms has not been previously validated.

Methodology
The authors propose a lightweight, end-to-end perception-and-control system utilizing a 77–81 GHz mmWave FMCW RADAR sensor on a small-scale quadrotor. The system architecture is divided into three main components:

Perception and Tracking:
- Detection: The system employs a sliding-window algorithm combined with DBSCAN clustering to process raw RADAR point clouds at 100 Hz. It distinguishes dynamic from static points using radial velocity thresholds, aggregating frames to generate 3D bounding boxes with state estimates $(x, y, z, v_x, v_y, v_z)$ .
- Tracking: A SAFE-Interacting Multiple Model (IMM) tracker is used to robustly track obstacles. This architecture bounds trajectory drift during model switches without high computational cost, utilizing a Global Nearest Neighbor (GNN) association and Kalman filters (constant velocity and constant acceleration).
- Latency Analysis: The authors derive analytical bounds for successful avoidance, defining a critical distance $d_{crit}$ based on the total end-to-end latency ( $t_{end} = t_{det} + t_{avoid}$ ), the evasion time ( $t_{evade}$ ), and the relative velocity ( $v_{rel}$ ). Successful avoidance requires triggering the maneuver before the obstacle enters $d_{crit}$ .
Control and Avoidance:
- The system utilizes a Distributionally Robust Acceleration Control Barrier Filter (DR-ACBF). Unlike heavy quadratic programming solvers, this controller uses a Gauss-Southwell deterministic projection to compute evasive accelerations within a linear half-space.
- To handle RADAR noise and estimation errors, the framework incorporates a Distributionally Robust Conditional Value-at-Risk (DR-CVaR) trigger. This mechanism adaptively expands safety margins based on worst-case probability distributions, ensuring robust collision avoidance even with noisy measurements.
Implementation:
- The system was implemented on a Raspberry Pi 4B (Pi4B) companion computer, running Ubuntu 20.04 and ROS2 Galactic, interfaced with a PX4 flight controller.
- The RADAR sensor is configured for high velocity resolution and a wide field of view, optimized for detecting small Radar Cross Section (RCS) objects.

Key Contributions

First mmWave RADAR-based Fast DOA: To the authors' knowledge, this is the first system to demonstrate fast onboard DOA using mmWave RADAR on a resource-constrained UAV without GPU acceleration.
Latency and Spatial Bounds Analysis: The paper provides a theoretical derivation of the temporal and spatial constraints required for successful DOA, relating sensing range, relative speed, and control delays.
System Validation: The authors validate the system on both a workstation and a Pi4B, demonstrating real-time feasibility with an end-to-end latency of approximately 14 ms.
Comprehensive Dataset: The work includes a dataset of 390 throws involving three different object sizes under varying visibility conditions (light, dark, and smoke).

Experimental Results
The system was evaluated through stationary hardware-in-the-loop (HiL) experiments and onboard flight tests:

Detection Accuracy: In light and dark conditions, the system achieved position errors of less than 0.15 m in the x-direction, 0.93 m in y, and 0.87 m in z. In smoke conditions, where ground truth was unavailable, the dispersion of detected positions remained consistent with light/dark results, indicating robustness.
Avoidance Performance: Across 390 throws in stationary experiments with varying object sizes (0.1 m to 0.3 m), the system achieved an overall avoidance rate of over 80% for distances greater than 1 meter. The DR-ACBF controller successfully triggered avoidance at an average distance ( $d_t$ ) of 2.12 m to 4.03 m, well beyond the calculated critical distance ( $d_{crit} \approx 1.02$ m for stationary tests).
Flight Experiments: In 20 onboard flights (hovering and forward flight at 1 m/s), the system demonstrated reliable avoidance at obstacle speeds exceeding 6 m/s. The end-to-end computation latency on the Pi4B was 13.65 ms, satisfying real-time requirements.
Failure Analysis: The system failed primarily when ultra-small, non-metallic objects (e.g., 0.1 m tennis balls) in smoke were detected too late (below $d_{crit}$ ) due to low RCS, or if CPU spikes increased latency beyond the safety margin.

Significance and Claims
The paper claims that mmWave RADAR is a viable primary sensing modality for fast DOA on small UAVs, offering a strong trade-off between robustness (in smoke and low light), latency, and onboard computational efficiency. The authors emphasize that their system achieves robust detection and avoidance without relying on deep learning or GPUs, making it suitable for deployment on resource-limited edge hardware. The results confirm that with appropriate latency management and robust control strategies (DR-ACBF), RADAR can support high-speed dynamic obstacle avoidance, a capability previously dominated by vision and LiDAR systems. Future work is noted to extend to faster-moving ego-UAVs in outdoor settings and multi-object avoidance.

More like this