#227 The ugly truth about LIDARs for obstacle avoidance (RPLIDAR A1, Arduino, ESP8266)

Obstacle avoidance is an essential topic for self-driving vehicles. In this discussion, we often hear the word “LIDAR.” These devices recently got cheaper and affordable for many Makers. But are they worth your money for obstacle avoidance?

Links:

My Bitcoin address: 19FSmqbBzb5zsYB1d8Bq4KbxVmezToDNTV
If you want to support the channel, please use the links below to start your shopping. No additional charges for you, but I get a commission (of your purchases the next 24 hours) to buy new stuff for the channel

Please do not try to Email me or invite me on LinkedIn. These communication channels are reserved for my primary job

[trailwander]

Andreas

It is not just this video. It is all your videos that are helping me learn more and more about electronics, robotics, etc. Thank you. This was another helpful video

[TimmmmCam]

TL;DR: A 2D lidar only detects obstacles in a 2D plane.

[thefatmoop]

Lidar is in fact very good at detecting obstacles and there's a major point missed here: autonomy systems don't just judge a single lidar scan, but many successive scans over time to answer the 'is there an obstacle here'. It also helps lidar tremendously when it's on a moving platform with varying roll and pitch and there's a good positioning system (RTK gps and high rate IMU) aiding in positioning-tagging points.

[DougHanchard]

Another great video!

One of the early systems I was involved in 2009 was using both LIDAR and Ultrasound together as you show in the last segment.

We had the same limitations you discovered. However, we also came up with a solution. Detect, query, target, verify.

The Lidar sensor would detect an object. If limited return data fell outside of programmed limits, the Ultrasound sensor was sent a command to target the unknown obstacle. Of course, if the sensor was fixed, this had limitations. Our solution, to mount the ultrasound sensor on a plate that had X and Y servo motors to control the directional plane required for targeting. This raised two more problems, time and number of queries per second (plus latency).

Our solution, install four ultrasound sensors, one at each corner of our robot.

Raspberry did not yet exist. We too had the exact same data processing issue. In fact we still had issues when we attempted to use Motorola RISC 88000 processors. In summary, we did not properly plan project objectives.

Our solutions were far too complex and over engineered.

We were trying to create high fidelity resolution that was never required and did not focus our efforts to a simple problem, target, define and avoid.

[ats]

Excellent video! For short range obstacle avoidance, an ultrasonic system with an omni source and a multi-element phased array receiver can be used for DoA estimation with full spatial coverage (a four element pyramid would be the minimal number of phones for full spacial coverage). The main problem with this approach is that it can only see one thing at a given two way range from the system at a time, and if there are multiple reflections received at the same time, the system can become confused, but it is generally possible to throw out bad results by requiring all returns to come in at consistent times at each receive array phone.

I was actually working on this problem tonight for an underwater military application when I took a break to watch your movie. But in my case, the problem is to detect marine mammals anywhere near a high-power low frequency acoustic surveillance array. During peacetime, these arrays must be turned off when marine mammals are nearby because having them wash up on a beach is terrible publicity for the Navy. The offshore oil exploration vessels have the same problem...

[NuclearWinter]

I have a better title: 2D LiDAR struggles when 3D is needed
There is no “ugly truth” here, just common sense it seems?

[Vatsek]

Sir, your tank has no weapons, you need to fix that problem ASAP.

[Mr]

I appreciate the objective manner in which he approaches the topics.

[roidroid]

Suggestion: Tilt the LIDAR down a few degrees at the front, and put a mirror behind it to reflect the backwards laser beams back towards the front again but on an UPWARDS tilt. This will effectively give you two forward facing detection planes, on different tilted angles, one up and one down. You just sacrifice your backwards facing detection.

[JayDAnderson]

The auto industry has solved this LiDAR blindness problem by using 75GHz RADAR to supplement the LiDAR sensor. Other types of sensors that are more nearsighted can be used to improve nearby object accuracy.

[sebbes]

6:45 You can put the lidar on a rotating wedge, in that way you can know at what angle the liar are (because you know the rotation of the wedge) and suddenly you have made a 3D lidar out of a 2D liar. (basically make the 2D "lazer disc" wobble around in 3D space, with whatever angle you need to cover the desired volume)
(A probable cause for problems is the G-forces created from the lidar's own rotation)

[pisoiorfan]

For slow moving vechicles a servo tilting the lidar would allow it to scan up/down a few degrees in the moving direction. Probably enough to detect descending stairs.

[adrianschneider]

What about letting the lidar tumble or nod in a controlled manner, eg. with a servo. So achieving a pseudo 3d operation. Just curious.

[klassichd]

Thanks for your contribution and the clarification.
But to use Swiss precision: Lidar (or Ladar) always means a time of flight method which the A1 do not use. So it is NOT a Lidar but a laser triangulation tystem originally developed for a vacuum cleaner robot.
Yes you arer right real and useful 3D Lidars are still very challenging and expensive. That is why the camera Systems got so popular in the automotive Industry over the last decade. Some car manufacturers even use a stereo camera to get a better estimation of real distances.

[lukdo]

Good Job! I like real information, verified behaviour. It is great for me to understand what to expect from these tools. Thank you for clear picture.

[alphonsesynrem]

As always Mr Andreas, very interesting topic. Thank you.

[FooBar]

what you're describing is the limitations of the cheap toy LIDAR systems; an industrial LIDAR system, used on cars, has none of those issues, it scans in 3D space and it does so very fast using an array of lasers and detectors

[unlost]

Good approach to the topic. Like ur video. Looking for more of your videos now.

[SianaGearz]

You have described a wall as a mirror reflection; however a wall is not a mirror, it's close to a diffuse surface, so the light gets thrown back in all directions regardless where source light came from. The strongest emission of light is along the surface normal and falls off from there on out towards no emission at 90° to the normal. The sensitivity of the sensor will also have a cutoff, which is why your LIDAR can't see these walls very far. See Lambertian BRDF.

[NatTuck]

For simple obstacle avoidance, a 2D lidar is great. Many environments are sufficiently constrained that planar sensing is fine. Once you start worrying about complex terrain it becomes a really hard problem. You discuss hanging obstacles - once you start getting into those, you might as well start worrying about *deformable* obstacles. The simplest example is long grass: any simple sensor will register it as an obstacle, but you want to drive right over it. At that point you need multiple sensors - probably a touch sensor combined with something like a planar lidar.


As for not wanting to process all the data from a planar lidar, just wait until you realize you want to add a camera. It's definitely worth having at least raspberry pi level processing on a mobile robot, if not a full on laptop-type computer.