Real-time obstacle avoidance is a tricky problem for small autonomous drones. Onboard computing power is limited, but the aircraft still needs to detect obstacles and react quickly enough to avoid them.
Depth cameras are one way to approach the problem. They provide detailed spatial information that can be used for obstacle detection and environmental perception. But acquiring depth data is only part of the process. The system also needs to process that information and turn it into flight control decisions.
On a small drone running perception algorithms on something like a Raspberry Pi, that processing overhead matters.
A recent Nature paper, Efficient robot navigation inspired by honeybee learning flights, includes an interesting change in the researchers' obstacle avoidance hardware.
The team initially used an Intel RealSense D435i depth camera to detect obstacles. In later experiments, they switched to three Benewake TF-Nova LiDAR rangefinders, positioned to monitor the left, center, and right directions ahead of the drone.
The paper primarily focuses on a navigation method inspired by honeybee learning flights, rather than a direct comparison between depth cameras and LiDAR. Still, the change raises an interesting question:
If an obstacle avoidance algorithm only needs to know whether something is blocking a particular direction, does it really need to process an entire depth image?
1. What Actually Changed When the Researchers Switched to LiDAR?
The researchers built a custom quadcopter using a Pixhawk 6C Mini running PX4 firmware for flight control, with a Raspberry Pi 4 handling onboard perception and higher-level navigation.
The platform also included an omnidirectional camera, an optical flow sensor, and a Benewake TFmini-S LiDAR rangefinder.
The TFmini-S provided altitude measurements to support motion estimation alongside optical flow and other sensor inputs. The omnidirectional camera supplied images for the visual homing system.
Obstacle avoidance was handled separately.
In the initial configuration, the researchers used an Intel RealSense D435i depth camera. The depth image was divided into three regions: left, center, and right.
The system calculated the average depth in each region to determine whether an obstacle was present and which direction appeared more open.
This is a fairly straightforward approach, but it still requires processing depth image data to extract the measurements used by the avoidance algorithm.
In later experiments, the team replaced this setup with three TF-Nova LiDAR rangefinders, each covering a 1° × 14° field of view.
One sensor monitored the left, another looked straight ahead, and the third covered the right.
Instead of extracting distances from different parts of a depth image, the avoidance algorithm could read directional distance measurements directly from the sensors.
That changes the type of information available to the controller.
A depth camera provides a much richer representation of the space ahead. Three directional LiDAR sensors provide considerably less information, but that may be sufficient when the algorithm only needs a few distance measurements.
Of course, this doesn't mean three LiDAR sensors are always more computationally efficient than a depth camera. Processing requirements depend on the camera hardware, how depth data is generated, the software implementation, and the overall system architecture.
In this particular reactive avoidance setup, however, direct range measurements eliminated the need to extract those distances from depth images.
2. How Can a Drone Avoid Obstacles with Just Three Distance Measurements?
The researchers used a finite-state machine (FSM) for reactive obstacle avoidance rather than building a detailed 3D map and planning a complete path around every obstacle.
The three TF-Nova sensors continuously measured distances in different directions ahead of the drone.
During normal flight, the system checked the forward distance against a predefined stopping threshold.
If an obstacle was detected within that threshold, the current navigation routine was interrupted and the avoidance controller took over.
The drone then used the left and right distance measurements to determine which direction offered a possible route around the obstacle.
For example, if the forward direction was blocked, the aircraft could rotate toward a clearer side and perform an avoidance maneuver.
If all three directions were blocked, the controller attempted a larger directional change to find another way through.
None of this required a complete 3D reconstruction of the surrounding environment.
The LiDAR sensors were not being used for mapping. Their job was to provide obstacle distance measurements that the controller could use to trigger specific avoidance actions.
There are obvious limitations to this approach.
Three directional sensors only cover a small portion of the space ahead. An obstacle located between the sensing regions could be missed entirely.
The stopping threshold also needs to be considered alongside the drone's actual flight dynamics.
At higher speeds, detecting an obstacle is not enough. The aircraft still needs sufficient time and distance to decelerate, stop, or change direction.
Sensor range, mounting angles, field of view, flight speed, and controller response time all affect how well the system performs.
For a relatively slow-moving drone operating in a controlled environment, directional range sensing may be sufficient for a simple avoidance strategy.
For faster flight or environments containing thin obstacles, dense vegetation, or irregular structures, three narrow sensing regions would leave significant gaps in coverage.
3. What Happens to Navigation After the Drone Avoids an Obstacle?
Another detail worth looking at is how the obstacle avoidance system interacts with the drone's higher-level navigation.
Bee-Nav was developed to investigate autonomous navigation and homing inspired by honeybee learning flights, without relying on a detailed global environment map.
The researchers implemented several flight phases, including learning flights, outbound navigation, inbound navigation, and final visual homing.
The basic obstacle detection and avoidance logic remained similar across these phases, but the way the drone resumed navigation after an avoidance maneuver was different.
During outbound and inbound navigation, the aircraft generally attempted to realign with its previous navigation target after avoiding an obstacle.
The goal was to continue the original route rather than treat obstacle avoidance as a separate mission.
During final visual homing, things worked differently.
Instead of simply returning to the previous heading, the drone captured a new image from its current position and passed it to the visual homing network to estimate an updated flight direction.
This matters because an avoidance maneuver can change both the drone's position and orientation.
During visual homing, the next movement needs to be determined from what the aircraft currently sees, not necessarily from the direction it was following before the obstacle appeared.
Obstacle detection and autonomous navigation are closely connected, but they solve different problems.
A LiDAR rangefinder can tell the controller that an obstacle is present at a certain distance. It does not determine where the drone ultimately needs to go.
In this architecture, the reactive controller deals with immediate obstacles, while the higher-level navigation algorithm decides how the aircraft should continue toward its destination.
4. Is This a Useful Approach for Small UAV Developers?
For developers working with Pixhawk, PX4, or Raspberry Pi-based autonomous drones, this setup offers a practical example of matching sensor inputs to the actual requirements of a control algorithm.
If an avoidance algorithm doesn't need a complete environment map, it's worth asking what information it really needs.
Consider a small experimental drone flying at relatively low speeds.
If the main requirement is to detect obstacles ahead and make simple left-or-right avoidance decisions, using direct range measurements may be easier to implement than building an entire depth-image processing pipeline.
But the requirements change when the aircraft needs to navigate through narrow openings, detect complex obstacle boundaries, or plan continuous trajectories at higher speeds.
In those situations, a few directional distance measurements may not provide enough information.
The sensor configuration used in the Bee-Nav study is also interesting because the researchers didn't try to make one sensor handle every perception task.
Different components served different purposes:
- Three Benewake TF-Nova LiDAR rangefinders: Measured obstacle distances in the left, center, and right directions for reactive avoidance.
- Benewake TFmini-S: Provided altitude measurements to support onboard motion estimation.
- PMW3901 optical flow sensor: Supplied observations of the drone's movement relative to the ground.
- IMU and extended Kalman filter (EKF): Combined sensor measurements to estimate the aircraft's motion and state.
- Omnidirectional camera: Captured visual information for the learned homing system.
The distinction between TFmini-S and TF-Nova is worth making. Both are LiDAR rangefinders, but they served different purposes in this experiment.
The TFmini-S was used for altitude measurement.
Optical flow provides information about apparent ground movement, but converting that image motion into useful horizontal velocity estimates requires knowledge of the drone's distance from the ground.
That's where altitude measurements become important.
The three TF-Nova sensors, meanwhile, were dedicated to detecting obstacles ahead of the aircraft.
Separating these functions makes it easier to understand what each sensor contributes and allows developers to work on individual parts of the perception system.
That doesn't necessarily make the overall hardware integration simpler.
Multiple independent rangefinders still introduce practical considerations such as mounting space, power distribution, communication interfaces, sensing coverage, and data synchronization.
Whether this approach makes sense depends on the aircraft and what it's expected to do.
5. Does Drone Obstacle Avoidance Really Need More Sensor Data?
Coming back to the original hardware change, replacing a depth camera with three directional LiDAR sensors doesn't suggest that depth perception is unnecessary.
Nor does it establish that LiDAR is inherently better than a camera-based solution.
The two approaches simply provide different kinds of information.
A depth camera is useful when an application needs a more detailed understanding of the surrounding space, such as analyzing obstacle geometry, identifying traversable regions, or supporting local path planning.
Directional LiDAR measurements are more limited, but also more direct.
If the control algorithm only needs to know whether the forward path is blocked and which side has more clearance, a handful of range measurements may be enough.
That's essentially the approach used in the Bee-Nav experiments.
The researchers didn't build a separate 3D map for obstacle avoidance. Instead, they used directional range measurements to trigger predefined control responses through a finite-state machine.
Visual homing, motion estimation, and obstacle detection were handled by different parts of the system.
For small UAVs, where payload capacity, onboard computing power, and energy consumption are all constrained, this is a useful design choice to examine.
The real question isn't how much environmental data a sensor can produce, but how much of that data the flight controller actually needs.
Of course, more demanding environments may justify wider sensing coverage, depth cameras, 3D LiDAR, or more sophisticated local planning algorithms.
The appropriate sensing architecture ultimately depends on the mission.
What I'd be interested to hear from other UAV developers is where you draw the line between simple range-based obstacle avoidance and more complete spatial perception.
If you're building a small autonomous drone, when would you choose multiple directional LiDAR rangefinders over a depth camera?
And which factor would influence your decision most: flight speed, obstacle density, or available onboard computing power?
Research Reference
Ou, D. et al. Efficient robot navigation inspired by honeybee learning flights. Nature 653, 1039–1046 (2026).
Comments