RADAR Perception for Dynamic Obstacle Avoidance onboard small-scale Quadrotor UAVs
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.
Rosa: I'm Rosa.
Dev: And I'm Dev, here with Taro.
Rosa: We're looking at the paper: "RADAR Perception for Dynamic Obstacle Avoidance onboard small-scale Quadrotor UAVs".
Dev: Here's the gist of the paper. Fast dynamic obstacle avoidance (DOA) on uncrewed aerial vehicles (UAVs) demands not only low-latency control and actuation but also reliable perception with sufficient sensing range for accurate obstacle detection and speed estimation <ref:2608.01855#pg2> This letter presents, to the best of our knowledge, the first mmWave RADAR-based perception-and-control system for fast onboard DOA <ref:2608.01855#pg2> Our system achieves position errors of less than 0.15 m, 0.93 m, and 0.87 m in x, y, and z directions for 300 experiments with three different object sizes and varying visibility (light and dark), and a similar spread for 90 experiments in smoke <ref:2608.01855#pg2> An onboard implementation on a Raspberry Pi 4B demonstrates real-time feasibility with an end-to-end sensing-to-command latency of approximately 14 ms <ref:2608.01855#pg2>.
Rosa: System Modeling and Latency Analysis The DOA timing can be partitioned into detection latency tdet, avoidance latency tavoid, and evasion latency tevade <ref:2608.01855#pg2>.
Dev: The resulting end-to-end processing latency is tend = tdet + tavoid <ref:2608.01855#pg2>.
Rosa: The critical distance dcrit defines the boundary for successful DOA <ref:2608.01855#pg2> Successful DOA not only requires sufficiently low latencies but also a robust perception of the spatial setting <ref:2608.01855#pg2>.
Paper summary: Rosa: So, this paper, "RADAR Perception for Dynamic Obstacle Avoidance onboard small-scale Quadrotor UAVs," it's about how you get fast obstacle avoidance working on these little drones. It points out that you need low latency control and really good sensing range to detect obstacles accurately and figure out their speed too.
Dev: Right, it claims this is the first mmWave RADAR based perception-and-control system for fast onboard dynamic obstacle avoidance on a quadrotor UAV. It also derives and analyzes how latency relates to sensing range, relative speed, and control delay, giving you the conditions needed for successful avoidance.
Taro: What matters here is that they're looking at position errors of less than zero point one five meters in the x, y, and z directions across three different object sizes and varying visibility conditions. And they also showed a similar spread when testing in smoke over ninety experiments.
Rosa: The implementation part is interesting because they show it running on a Raspberry Pi 4B companion computer to prove it's feasible for real-time use <ref:2608.01855#pg1,on a Raspberry Pi 4B>. They report an end-to-end sensing-to-command latency of about fourteen milliseconds.
Dev: That latency figure, fourteen milliseconds, that’s what we need to worry about for a real control loop. The paper breaks down the total timing into detection latency, avoidance latency, and evasion latency. And it defines a critical distance dcrit which sets the boundary for successful dynamic obstacle avoidance on the system.
Taro: So they calculate this critical distance based on the relative velocity between the object and the UAV over that period of evasion time and detection time. That's a key part of their analysis showing what successful avoidance really means in terms of speed and timing.
Rosa: The actual detection model they built involves clustering, candidate updates, creating bounding boxes, and then using a tracker latency for perception. They used this SAFE IMM tracking architecture because it helps bound trajectory drift during model switches without taking too much computational power.
Dev: And for the avoidance side, they use a DR ACBF avoidance model that employs a Gauss Southwell deterministic projection to find the closest safe acceleration inside a linear half-space. That's how they translate the perception into actual evasive commands.
Taro: They validated this system on two platforms, including hardware in the loop for stationary experiments and an onboard experiment using that Raspberry Pi 4B companion computer <ref:2608.01855#pg1>. And they used three transmitters and four receivers for their preferred RADAR configuration, which gave them a maximum radial velocity of about plus or minus sixteen point two meters per second.
Rosa: The experimental results show clear differences in how the UAV avoids obstacles depending on whether it's hovering or flying forward, and whether it's in light or dark environments. They showed the DR ACBF avoidance works reliably, achieving an overall average of eighty nine percent success when the object ball velocity was over three point one meters per second and a maximum of eight point four meters per second.
Dev: But they also flagged some failure modes, which is important for a safety case. They explicitly state that their framework can fail if the detection time is less than or equal to the critical distance. This happens in two ways: late detection of ultra small objects, like targets under zero point one meter non-metallic in smoke, or CPU spikes which they say give total latency averages of thirteen point six five milliseconds.
Taro: So their overall success rate is more than eighty percent when dealing with a distance greater than one meter for three balls. And they show that the DR ACBF controller can enable fast, safe reactions in sixty to one hundred milliseconds while keeping the minimum separation margins above the critical safety distance.
Rosa: So, putting it together, this paper presents a lightweight mmWave RADAR-based perception and control system designed for real-time dynamic obstacle avoidance on resource constrained UAVs. It confirms that mmWave RADAR is a viable primary sensing modality for fast dynamic obstacle avoidance systems, offering a trade off among robustness, latency, and onboard computational efficiency.
Dev: The DR ACBF controller allows for quick, safe reactions in sixty to one hundred milliseconds while maintaining minimum separation margins above the critical safety distance. This means the system has a defined safety boundary based on the relative velocity and timing calculations.
Taro: This work is about making sure that when the world misbehaves with moving objects, the system doesn't just react quickly but reacts safely according to these derived constraints. It shows how you can design a controller that respects those safety bounds even when things are getting fast.
Rosa: For someone who only listens to this show, it means that for small drones flying in complex environments, using mmWave RADAR as your main way to perceive moving obstacles is a strong choice because it balances speed and the processing power you have on board. It proves that real-time avoidance is achievable with current onboard computing setups.
Conclusion: Rosa: So we've seen how they built this whole system, from using mmWave RADAR to get around obstacles on those little quadrotors in real life. Now we need to talk about what they actually achieved with this paper, "RADAR Perception for Dynamic Obstacle Avoidance onboard small-scale Quadrotor UAVs."
Dev: Yeah, the main thing is that they show a complete system that works without needing some super complex processing power on the drone itself. They're focusing on getting fast obstacle avoidance using this RADAR sensing method.
Taro: What I find interesting about it is how they handled those real-world failures. It wasn't just a perfect simulation; they tested it with three different object sizes and even in smoke, which means you can see if the system breaks when things get messy.
Rosa: And the results show pretty good accuracy for position estimation, under zero point one five meters in all directions for their experiments. That’s a solid number when you’re trying to navigate close to things.
Dev: It also shows they can run this on a Raspberry Pi 4B, and the latency they got from sensing to sending a command was about fourteen milliseconds, which is pretty fast for real-time control loops <ref:2608.01855#pg1,on a Raspberry Pi 4B>.
Taro: But remember, it's not all perfect either. They laid out exactly when the system might fail—like if an object is too small or if the computer gets overwhelmed by spikes in processing time.
Rosa: That’s the caveat, right? The paper really proves that mmWave RADAR isn't just a cool piece of tech for a lab; it can be a primary way to see what's around you when you’re actually flying.
Dev: It suggests that for small drones needing fast reactions, this trade-off between how robust the sensor is and how little processing power you need is pretty favorable.
Taro: It opens up possibilities for autonomous systems that need to react quickly in unpredictable environments, whether it's a delivery drone or something else entirely.
Rosa: And that’s where we’re going to look next—we’ll see how this kind of sensing capability could actually change what these drones can do on the ground.
Dnyandeep Mandaokar, Bernhard Rinner
University of Klagenfurt
cs.RO, cs.SY, eess.SY
Submitted: 2026-08-03
Updated: 2026-11-06
Comments: Preprint. This work has been submitted for possible publication and is currently under review
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 79/100
The gist: Fast dynamic obstacle avoidance (DOA) on uncrewed aerial vehicles (UAVs) demands not only low-latency control and actuation but also reliable perception with sufficient sensing range for accurate…
Key concepts
- End-to-end Processing Latency (tend)
- This measures the total time from sensing an obstacle to issuing a control command. It is broken down into detection, avoidance, and evasion latencies. The goal is to keep this latency low enough for the UAV to react quickly and safely during dynamic maneuvers.
- Critical Distance (dcrit)
- This defines the minimum safe distance required between the UAV and an obstacle for successful avoidance. It is calculated based on relative velocity over specific time periods, ensuring that the UAV has enough space to maneuver laterally away from a detected object.
Terminology
Summary
Fast dynamic obstacle avoidance (DOA) on uncrewed aerial vehicles (UAVs) demands not only low-latency control and actuation but also reliable perception with sufficient sensing range for accurate obstacle detection and speed estimation <ref:2608. p. 0 1
RADAR Perception for Dynamic Obstacle Avoidance
onboard small-scale Quadrotor UAVs
Dnyandeep Mandaokar1
and Bernhard Rinner1
Abstract— Fast dynamic obstacle avoidance (DOA) on uncrewed aerial vehicles (UAVs) demands not only low-latency
control and actuation but also reliable perception with sufficient sensing range for accurate obstacle detection and speed
estimation. This letter presents, to the best of our knowledge, the
first mmWave RADAR-based perception-and-control system for
fast onboard DOA. 1
RADAR Perception for Dynamic Obstacle Avoidance
onboard small-scale Quadrotor UAVs
Dnyandeep Mandaokar1
and Bernhard Rinner1
Abstract— Fast dynamic obstacle avoidance (DOA) on uncrewed aerial vehicles (UAVs) demands not only low-latency
control and actuation but also reliable perception with sufficient sensing range for accurate obstacle detection and speed
estimation. This letter presents, to the best of our knowledge, the
first mmWave RADAR-based perception-and-control system for
fast onboard DOA. Our system adopts a lightweight tracker based on interacting
multiple models and a controller based on control-barrier
functions that directly outputs evasive accelerations. It achieves
position errors of less than 0. 93 m, and 0. 87 m in x,
y, and z directions for 300 experiments with three different
object sizes and varying visibility (light and dark), and a
similar spread for 90 experiments in smoke. 87 m in x,
y, and z directions for 300 experiments with three different
object sizes and varying visibility (light and dark), and a
similar spread for 90 experiments in smoke. An onboard
implementation on a Raspberry Pi 4B demonstrates real-time
feasibility with an end-to-end sensing-to-command latency of
approximately 14 ms.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
System Modeling and Latency Analysis
The DOA timing can be partitioned into detection latency tdet, avoidance latency tavoid, and evasion latency tevade <ref:2608.01855#pg2>. The resulting end-to-end processing latency is tend = tdet + tavoid <ref:2608. p. 0 tavoid is the aggregate execution
time of the avoidance modules. The resulting end-to-end
processing latency is tend = tdet + tavoid. 2
critical distance dcrit defines the boundary for successful
DOA. If avoidance is triggered before the obstacle enters
dcrit, the UAV can still evade. dcrit is calculated based on
tend and the relative speed vrel of the obstacle wrt. Successful DOA not only requires sufficiently low latencies
but also a robust perception of the spatial setting.Where the paper says it
Where the paper says it
Where the paper says it
The analysis involves estimating tend for the preferred RADAR sampling rate and calculating dcrit based on relative velocity vrel <ref:2608. p. 2 In order to assess the perception
limits for DOA, we estimate tend for the preferred RADAR
sampling rate. dcrit is then estimated based on tend and the
assumed relative obstacle velocity vrel. For the detection range R and lateral acceleration budget a,
we compute the allowable UAV speed as
vmax(R, a, tend) = max
0,
√
2aR − a · tend
. (1)
The formulation of the critical distance dcrit is based on
the relative velocity vrel over the period of tevade and tend. (1)
The formulation of the critical distance dcrit is based on
the relative velocity vrel over the period of tevade and tend. The successful avoidance implies that the UAV must laterally
displace itself by at least the radius of the UAV and the
obstacle, i. Copyright may be transferred without notice. arXiv:2608.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Implementation Details
The implemented detection model consists of clustering, candidate update, bounding box creation, and tracker latency <ref:2608.01855#pg2>. The SAFE-IMM tracking architecture is utilized for RADAR perception due to its ability to bound trajectory drift during model switches without incurring high computational costs <ref:2608. p. 3 Tracking
We utilize a SAFE-IMM tracking architecture [15]. The
choice to utilize this architecture for RADAR perception
is primarily driven by its ability to bound trajectory drift
during model switches without incurring high computational costs. Avoidance
This work implements a DR-ACBF avoidance model, a
lightweight controller producing an acceleration avoidance
command. Instead of solving a heavy quadratic program,
the DR-ACBF leverages a Gauss-Southwell deterministic
projection [14] to find the closest safe acceleration inside
a linear half-space.Where the paper says it
Where the paper says it
The system is validated on two platforms, including a hardware-in-the-loop (HiL) stationary experiment and an onboard experiment using a Raspberry Pi 4B companion computer <ref:2608. p. 3 The DR-ACBF is preferred due to its
suitability for noisy RADAR detection, which is accounted
for by the ACBF’s effective radius and safety margins. We validate the system on two computer platforms: a
hardware-in-the-loop (HiL) stationary experiment on a laptop
and a flying experiment with a Raspberry Pi4B companion
computer. The PX4 is configured
for a maximum allowable acceleration a =15 m/s2
. Sensor Configuration
The proposed RADAR profile is configured for frame
settings of 3 TX (Ntx) and 4 RX (Nrx) for improved
velocity resolution, carrier frequency fc =77 GHz, and 110
loops, giving 330 chirps per frame with 110 Doppler bins
(ND) per TX channel. Copyright may be transferred without notice. arXiv:2608.Where the paper says it
Where the paper says it
Where the paper says it
Experimental Results
The experimental validation is performed in two settings: a stationary HiL experiment and an onboard experiment using a Pi4B <ref:2608.01855#pg2>. For the onboard experiment, the avoidance is primarily focused on a single object to assess the avoidance latency of RADAR perception <ref:2608. p. 4 The experimental
validation is performed in two settings: a stationary HiL
experiment to assess detection capabilities, with 390 throws
across three ball sizes, and an onboard experiment validation
using a Pi4B as the companion computer with 20 flights. For the onboard experiment, the avoidance is primarily
focused on a single object to assess the avoidance latency
of RADAR perception.Where the paper says it
The results demonstrate clear differences in UAV avoidance performance between hovering and forward flight in light and dark environments <ref:2608. p. 6 The smoke scene was excluded due to
MOCAP failure. The results demonstrate clear differences in UAV avoidance performance between hovering and forward flight in
light and dark environments (Table V). Instead, we discuss the average and
maximum velocity to define the operational baseline and the
system’s verified upper performance bounds. The DR-ACBF avoidance shows reliable avoidance, with
an overall average of 89% at an average ball velocity of over
3. Following the setting in Figure 1, DOA challenges
can be summarized into detection latency, detection range,
avoid latency, and evade latency. The detection range R
specifies the maximum distance at which obstacles can
be robustly detected and influences the time available to
decide whether to trigger an avoid command. Copyright may be transferred without notice. arXiv:2608.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Safety and Failure Analysis
The formulation in Eq. 2 explicitly identifies the two primary conditions under which our mmWave RADAR framework can fail (dt ≤ dcrit) <ref:2608. p. 7 2) Failure Conditions: Despite robust design, the formulation in Eq. 2 explicitly identifies the two primary conditions
under which our mmWave RADAR framework can fail
(dt ≤ dcrit). 2 explicitly identifies the two primary conditions
under which our mmWave RADAR framework can fail
(dt ≤ dcrit). First, there is late detection of ultra-small
objects (e. First, there is late detection of ultra-small
objects (e. 1 m non-metallic targets in smoke). 04 m), a collision occurs. Second, CPU
spikes, which depend on the computing board and algorithmic efficiency, give tend with averages 13. Copyright may be transferred without notice. arXiv:2608.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
The overall detection and avoidance rate is more than 80% with a distance of more than 1 m for three balls <ref:2608. p. 5 1 m tennis ball, as its avoidance rate
is 70%. The overall detection and avoidance rate is more
than 80% with a distance of more than 1 m for three balls. Similarly, onboard experiments demonstrated reliable
avoidance performance under both dark and light conditions
at obstacle speeds greater than 6 m/s. The DR-ACBF controller enables fast, safe reactions within 60–100 ms while
maintaining minimum separation margins above the critical
safety distance. Similarly, onboard experiments demonstrated reliable
avoidance performance under both dark and light conditions
at obstacle speeds greater than 6 m/s. The DR-ACBF controller enables fast, safe reactions within 60–100 ms while
maintaining minimum separation margins above the critical
safety distance.Where the paper says it
Where the paper says it
Where the paper says it
Conclusion
This letter presents a lightweight mmWave RADAR-based perception and control system for real-time DOA on resource-constrained UAVs <ref:2608.01855#pg2>. Our results confirm that mmWave RADAR is a viable primary sensing modality for fast DOA UAVs, offering a strong trade-off among robustness, latency, and onboard computational efficiency <ref:2608. p. 7 The DR-ACBF controller enables fast, safe reactions within 60–100 ms while
maintaining minimum separation margins above the critical
safety distance. Our results confirm that mmWave RADAR
is a viable primary sensing modality for fast DOA UAVs,
offering a strong trade-off among robustness, latency, and
onboard computational efficiency. Finally, tevade
represents the time required for the UAV to overcome its
inertia and achieve the lateral displacement necessary for
collision avoidance; it is therefore not included in tend. The
This work has been submitted for publication.Where the paper says it
Where the paper says it
REFERENCES
arXiv:2608.01855v1 [cs. p. 0 arXiv:2608. 01855v1 [cs. 01855v1 [cs. RO] 3 Aug 2026Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improved perception latency management by utilizing a lightweight tracker based on interacting multiple models to reduce tdet, as described in Section I and Section IV-C, allowing for faster obstacle detection than traditional methods.
-
Enhanced safety margin calculation by using the DR-CVaR early-warning margins to dynamically adjust safety margins against worst-case probability distributions when evaluating the DRACBF avoidance trigger distance dt, ensuring robust collision avoidance under noisy RADAR detection.
-
Optimized onboard computational efficiency by deploying the system on resource-limited hardware like a Raspberry Pi 4B, achieving an end-to-end sensing-to-command latency of approximately 14 ms even when clustering dominates computation (Table VI).
-
Improved velocity estimation accuracy by configuring the mmWave RADAR profile with 3 TX and 4 RX for improved velocity resolution, yielding a maximum radial velocity of about ±16.2 m/s, which aids in more accurate relative speed calculation.
-
Enhanced avoidance decision making by ensuring that the trigger distance dt is always greater than dcrit based on the inequality dt > dcrit, where dcrit is calculated as vrel (tend + tevade), providing a mathematically verifiable condition for successful evasion.
Abstract
Fast dynamic obstacle avoidance (DOA) on uncrewed aerial vehicles (UAVs) demands not only low-latency control and actuation but also reliable, long-range perception to detect obstacles and estimate their motion in time to maneuver. We present, to the best of our knowledge, the first mmWave RADAR-based perception-and-control system for fast onboard DOA. We derive and analyze latency and spatial bounds that relate sensing range, relative speed, and control delay, yielding sufficient conditions for successful avoidance. We further extend prior latency analyses to include the decision latency of the tracker and trigger to evaluate them per trial. Our computationally efficient system combines a lightweight interacting multiple models (IMM) tracker with a control-barrier functions (CBF) controller that directly outputs evasive accelerations. In 300 throw experiments spanning three object sizes and varying visibility (light and dark), it achieves mean position errors of at most 0.15 m, 0.93m, and 0.87m in the x, y, and z directions, with a similar spread in 90 experiments in smoke. An onboard implementation on a Raspberry Pi 4B demonstrates real-time feasibility with an end-to-end processing latency of approximately 14ms per RADAR frame and a system latency of 75ms from first detection to the avoidance command. The UAV successfully avoided obstacles in all 20 tethered indoor flights and all 39 untethered outdoor avoidance maneuvers across three ball sizes, during both hover and forward flights at up to 5m/s with measured closing speeds up to 14.6m/s.
Related papers
- Rethinking Actions: Continuity-Aware Priors for World-Model Planning
- MOSAIC: Modular Obstacle and Spatial-Aware Framework for Manipulation In Cluttered Scenes
- SAGAS: Semantic-Aware Graph-Assisted Stitching for Offline Temporal Logic Planning
- DexSeed: Expanding Human Hand-Object Interactions into Robot Demonstrations for Dexterous Policy Learning
- LaST0.5: A Generalist World Action Model with Interaction-Centric Latent Reasoning
- CIALLO: Learning Shared Intent Representations for Cross-Embodiment Manipulation