Which is better, installing two LiDARs or one? Let's speak with data
A customer is developing a park patrol robot. Initially, the plan was to install one M360 on the roof. After testing the prototype, it was found that there was a blind spot: the blind spot at the rear of the vehicle was too large.
The M360 performs 360° horizontal scanning, theoretically covering all directions up, down, left, and right, with no "directional blind spot." However, in actual use, "being able to scan" and "being able to scan well" are two different things.
Why does even one radar have a blind spot?
The M360 is installed in the center of the roof, with the optical center facing up 5° (to better scan the ground). At this point, the ground coverage in the front is the best—0.05m blind spot + 70° vertical field of view, covering from a few meters in front to several tens of meters.
However, the situation is different at the rear. The robot's body itself blocks some point clouds—the shadow areas of the battery compartment, radiator, and antenna. Although the M360 performs 360° horizontal scanning, the points that hit the body itself do not return (either absorbed or reflected in the wrong direction), and these points are missing in these directions.
Measured data: On our AGV (body size 800×600×400mm), the point cloud density in the rear direction is about 40% lower than that in the front. In other words, the detection distance for obstacles at the rear may be much shorter than that at the front—the front can stably detect obstacles 30m away, while the rear may only be 15~18m.
For general applications, a 15m rear detection distance is sufficient. However, for high-speed scenarios (>1.5m/s) or dense scenarios (parking lots, logistics parks), a 15m detection distance may not be enough—the safety distance from detection to braking is insufficient.
The two LiDARs solution
The customer's solution is simple: one M360 at the front and one at the rear.
Installation location: The front LiDAR is installed facing forward, and the rear LiDAR is installed facing backward, both slightly tilted up 5°. The two LiDARs are connected to the industrial computer via an Ethernet switch and each has an independent IP address.
Point cloud fusion method: In ROS2, two livox_ros_driver2 nodes are used to receive data from the two LiDARs separately, and then pointcloud_to_laserscan is used to convert the 2D costmap for use by the navigation stack. It is also possible to use pcl_ros to merge the two point clouds to generate a complete 3D environment.
Measured results: The point cloud density in the rear direction has recovered from the original 60% to over 95% (the remaining 5% loss is unavoidable due to the body structure). The detection distance at the rear has increased from 15m to 30m, basically consistent with the front.
Comparison of Three Fusion Solutions
We tested three methods of point cloud fusion:
Solution 1: Independent topics, processed separately.The point clouds from the two radars are published as independent ROS2 topics, and the downstream obstacle avoidance nodes subscribe to the two topics respectively, making obstacle avoidance judgments individually.
Advantages: Simple, no need to modify existing algorithms. Disadvantages: Duplicate detection may occur in the overlapping area covered by two radars—the same obstacle is seen by both radars, and the obstacle avoidance algorithm may not know how to handle it, possibly leading to overreaction.
Solution 2: Point cloud stitching, processed uniformly.Combine the point clouds from the two radars into a large point cloud (by transforming coordinates to the vehicle coordinate system), and then send it as a topic to the downstream algorithms.
Advantages: Downstream algorithms only need to process one point cloud, with simple logic. Disadvantages: Duplicates in the overlapping area need to be handled, otherwise, uneven point cloud density may affect algorithm performance.
Solution 3: Build maps separately, share the map.Each radar runs SLAM (each with a single FAST-LIO2 instance), and the sub-maps built are merged into a global map through a map fusion algorithm.
Advantages: Best robustness—one radar is obstructed, the SLAM of the other radar can still continue to work. Disadvantages: High computational load, requires two SLAM instances to run.
In actual projects, Solution 2 is used the most.. The point cloud stitching calculation is minimal (tens of milliseconds), and the deduplication of overlapping areas can be achieved with voxel filtering, which is also not complex to implement.
The installation spacing between two radars
This is often overlooked by many. How does the distance between two radars affect the system?
Anti-interference.The M360 features an active anti-crosstalk design, but if two radars are too close to each other (<20cm), the intersection angle between the laser beams is too small, which may reduce the effectiveness of anti-crosstalk.
Point cloud overlap.The distance between the two radars determines the size of the overlapping area. The greater the distance, the larger the overlapping area, but the point cloud density becomes less uniform (dense in the overlapping area, sparse in the non-overlapping area).
Our recommendation is: if the vehicle body allows, maintain a spacing of 60~100cm between the two radars. At this spacing, the anti-interference effect is the best, and the point cloud density in the overlapping area is also reasonable (not too dense or too sparse).
If the vehicle body is too small (<80cm), and the spacing is insufficient, you can reduce the scanning frequency of one radar to reduce interference. For example, the radar at the front of the vehicle runs at full speed, while the radar at the rear is reduced to half the frequency. Sacrificing a bit of frame rate for better anti-interference effects.
When is it necessary to use three or more
Most projects require only two radars. Situations where three or more are needed generally include:
Large vehicle body——For heavy-duty AGVs with a long wheelbase, one radar at the front and one at the rear, with an additional one in the middle to fill the side blind zone.
Need for all-around close-range perception——For example, a warehouse robot navigating through a dense shelf area needs to detect obstacles within 10cm in all 360° directions.
Quadruped robots/humanoid robots——The body attitude changes significantly, requiring multi-directional compensation.
The complexity of fusion with three or more radars increases significantly. It's not just about the point cloud stitching; there's also the allocation of computational resources – each radar's driver node, point cloud preprocessing, and SLAM instance all require CPU and memory.
We have tested four M360s running simultaneously on the Jetson AGX Orin (FAST-LIO2 × 4), with CPU usage around 85% and 8GB of memory. It works, but there's not much room left. If additional algorithms (object detection, path planning, etc.) are added, it might not be sufficient.
The solution with two radars requires much less computational power. Running two FAST-LIO2s on the Jetson Orin Nano, the CPU usage is around 40%, leaving room to run other algorithms.
Cost and benefit
The last question is whether it's worth it. The price of an M360 is not cheap, and adding another one adds to the cost.
But let's do the math: if a collision occurs due to the blind zone at the rear of the vehicle, the repair costs (radar bracket deformation, paint scratch, downtime loss) far exceed the price of a radar. For commercial products, reliability is more important than cost.
However, it also depends on the scenario. If it's a low-speed drive in an open outdoor environment, the collision risk is low, and one radar is sufficient. If it's a high-speed drive in a dense indoor environment, the safety margin provided by two radars is worth the investment.
The above is based on actual project testing; the M360 specifications are to be taken from the official TanTu ZhiXing.