Optical Cliff Sensor Limitations

[glnc222]

With certain black floor materials carpet and hard floor, the drop or cliff sensors on vacuums detect false drops and fail to clean. An experiment was performed with a possible alternative to the standard proximity sensors which use changing intensity of IR light reflection with distance.

[edit] Maybe robots are just using the wrong parts. One sensor 22-70mm from Omron described as "convergent reflective" specifies
"Robust performance with little influence from sensing object color or material.
Detects objects with mirror reflective surfaces and transparent objects."
Omron B5W-LA01 data sheet https://www.omron.com/ecb/products/pdf/en-b5w_la01.pdf
Actually cheaper than the Neato Robotics Sharp sensor because lacks processing of the pulse signal, and just includes a power transistor for the emitter and output from the detector -- four wire instead of three.
Description at https://www.components.omron.com/produc ... la01/index
With spec for a particular distance, looks like the parallax method described in this post below.
(the only other model seems to be for 5mm, too close for cliffs)

[edit] A post in the Neato forum describes the Sharp Co. proximity sensors used as working by a triangulation method, similar to the short range lidars, parallax range finders. One reference so describing the Sharp components is https://wiki.eprolabs.com/index.php?tit ... GP2YOA21YK
Yet if they really worked this way I do not see why they would be so sensitive to object surface quality.
A detail in the Sharp description above might be relevant: instead of a digital camera sensor with separate pixels (as used in the lidar I think, measuring both angle and intensity of light return), the sensor just outputs an analog voltage varied with position, processed for output (these sensors also pulse the light to discriminate ambient light conditions etc.). Perhaps that sensor output is also influenced by the light intensity. These sensors measure distance while I think the Omron one just returns a binary indicator of objects in range.



In this test a parallax method was used less dependent on the intensity of reflection to measure distance. When the emitter and detector spreads cross there should be no reflection detected. The standard method relies upon the narrower spread of the focus of either detector or emitter resulting in less intense reflection over distance with beam spread, allowing calibration of distance. This method is more vulnerable to distortion from variations in floor material properties.

[edit] Note that if intensity variation from material property exceeds that of distance variation, increasing sensitivity to see black material would ruin normal material distance measurements. Only a parallax method could use higher sensitivity.

A fairly high powered emitter tightly focussed was used. [edit] Apparently not compared to the over-driven TV remote methods mentioned later below are used, with short duty cycles up to half amp. What is in the standard proximity components is not clear, trade secrets not in data sheets.
[edit] A Sharp sensor advises a large 10uf capacitor 5v power connection support, maybe for the emitter pulses.

The components worked well on light colored material and some black materials such as luggage nylon, and borderline on a black textured vinyl cell phone cover, still workable.
However, black knit cloth completely defeats detection, at least in IR. This probably means optical methods of cliff detection will not work on the problem floors reported, even with enhancements. Sonar would be needed or some mechanical feeler. Unless there are UV emitters and detectors? [edit] Some higher power emitter and other strategies mentioned later below might still make optical methods usable.
[edit] Samsung Powerbot sensors detected black knit cloth, so the problem material remains hard to find.

[edit] More sensitive TV remote IR receivers do work on the black knit fabric. See later post.

I did not have an IR laser to test, just a laser level in visible light not detected in IR. Powerful lasers present safety hazards, with lidars low powered accordingly so probably not useful.

[edit] PC laser mouse works on the black fabric. See later post.

Perhaps some camera sensor would be more sensitive than single phototransistors but likely too expensive. Such sensors are used in the vacuum lidars employing parallax for full range finder operation, where the return is indeed weak, so maybe in principle they would work aside from the economics. Yet with high enough sensitivity, emitter reflections between the floor and bottom of the robot could present a problem achieving the necessary discrimination.

Since vacuums move on rollers positioning the cleaning head and robot stance, I wonder if some mechanical sensing could be attached to those. However, they cannot easily distinguish navigable short drops from dangerous cliffs.

Samsung precautions in their user guides the robots will not work on black floors. At least they are up front with disclosure.

The emitter shows a useful focus in the picture but the detector, specified 20 degrees, was picking up returns far outside the supposed focus. There was still a marked voltage variation of 2v in the output over a half inch range at the cone crossing. Background without the emitter output virtually zero, along with any black knit cloth reflection. An additional lens would be needed to get proper detector focus.

[glnc222]

Correction: the black knit fabric defeating the phototransistor still works with LogiTech mouse for the PC. So maybe those lasers could be used for more effective cliff sensors.

[edit] The Roomba 980 has a mouse variant in it for distinguishing carpet from hard floors to optimize suction power. That feature is appearing on several brands now.

[mfortuna]

I know it is a throwback but two front mounted drop switches could do the job. They would need rollers and spring mounting but it seems doable. Yet another thing to get dirty I suppose.

I did a create project that uses an ultrasonic distance sensor. That could be another solution. Instead of pinging forward you ping down. Sounds like a potential project.

[glnc222]

I am getting the impression the common phototransistor proximity sensor was used for cliff sensors on robots because cheaper than contruction of mechanical probes, optical more reliable, simple mounting etc., maybe more durable. All at the expense of needed sensitivity. Going to laser tech to improve that, maybe the economy of simple mechanics would resume its proper place. Vacuum robots have gotten more expensive as they get better, and the cheaper ones now are getting a lot of attention -- are all those expensive advances worthwhile?

The mouse technology is much more sensitive, see Wikipedia https://en.wikipedia.org/wiki/Optical_mouse
They detect texture of surfaces with video camera type sensors, cheapened by mass production of dedicated specialty IC's with intergrated image processing -- some even including the laser. They do sit much closer to the surface than a cliff sensor, though, don't know what that implies. Logitech even developed a dual laser "darkfield" mouse capable of reading glass table tops as those became more popular in decor.

As for detecting differences between carpets and hard floors, I am not sure all brands are using mouse tech like Roomba. Differences in drag on the brush and drive wheels might be an indicator detectable in motor currents, and much less expensive. I don't see any special detectors on the bottom of the Samsung vacuums. Neato Robotics once mentioned how run time was reduced by introducing a better brush which happened to have more drag, and carpet run time was always substantially less than hard floors.

Cliff sensors require some smarts in the interpreting software. The sensors must get tripped whenever the robots climb obstacles needing combination with other inputs, perhaps from the accelerometer IC's common in these devices, as well as extension switch sensors on wheel suspensions.

[glnc222]

Encoded TV IR remote receivers also seem more sensitive than simple phototransistors, also made with specialty IC's mass produced like mice. A DIY simple proximity detector using one TSOP1 for Arduino's etc. appears in one article. They have wide angle lenses for convenient remote controls, so an additional focusing lens might make them practical for parallax cliff detection with a simple narrow beam emitter. The emitter need only be pulsed with an oscillator like a 555 IC at the standard 38khz used, and oscillation is needed to exclude background illumination anyway.

Some of these receivers tune to specific standard digital encoding standards besides mere carrier frequency, but some are simpler and leave that to the external processing.

[edit] Trying an old Radio Shack receiver and key chain remote tomorrow, on black cloth...

[glnc222]

TV Remote Receiver Test -- higher sensitivity

The TV remote receiver works on the black knit fabric defeating simple phototransistors. Tested with a Radio Shack 38khz receiver (actual mfg unknown) and cable box remote repeating a button.

The receiver specifies 45 degree lensing. So add a lens to tighten focus as needed for parallax drop sensors.
An article mentions TV remotes use high powered short duty cycle emitter power up to 500ma pulse peak current, because of the short cycles. That is how they work across long distances.
A ready made proximity sensor using such receivers in the standard fashion is British
https://www.floris.cc/shop/en/pololu-br ... esults=164
high intensity for up to 24 inches out to target. Low intensity versions also supplied. Mentioned just as illustration.
Using 555 oscillator for 38khz emitter drive described at http://www.robotroom.com/Infrared555.html. Too avoid adjusting a trimmer for accurate tuning a microcontroller might be used, or some specialized part for these remotes.

I previously used these with a key chain remote to make a virtual wall system for Neato robots equipped for mag strip boundary markers (thread "optical boundary marking" in Neato forum).

So just suitable focus of emitter and receiver might well produce a parallax drop sensor for dark floors, presumably less expensive than laser technology from mouse pointers. Not my specialty.

[glnc222]

More Techniques

1. Oscillating signal detection example. Detects TV remotes without decoding, not exact frequency dependent so fixed component 555 oscillators usable.
The long exposure technique from astronomy can accumulate weak signals over time for higher sensitivity.
Narrow focus phototransistors vs wider angle integrated TV remote receivers.
Works on problem black knit fabric with TV remote test.

2. Many DIY detectors employ multiple emitters for higher power, besides high power short duty cycle in TV remotes.
Usable with parallax method for cliff detection.

[glnc222]

Neato Sensor Test

Neato Robotics has used a Sharp short distance proximity sensor part no. Sharp GP2Y0A51SK0F in their vacuums. The same sensor is used for cliff detection and sideways wall detection, to follow walls etc.
Data sheet http://www.sharp-world.com/products/dev ... 51sk_e.pdf
Digi-Key.com part# 425-2854-ND

It has about 2.5v output range showing 1.8v on light colored material 1/2 inch away similar to a floor under the robot.
It was not affected by most black materials tried, suggesting it might over-drive the emitter in short duty cycles similar to TV remotes, but was not measured in that detail. They specify a large 10uF filter capacitor on the 5v power supply to it, probably to support the emitter pulsing load.

However, a brown sock material showed only 1.6v at half an inch, similar to the light material 3/4 inch away, a 50 per cent distance error. These devices are not the best for making the fine distinction between a drop which is navigable and one which is a genuine cliff to avoid, given the variation in material reflections, relative to distance variation.
I do not have the particular floor and carpet pattern material reported a problem in robot posts (Ikea being a three hour drive), but it looks like a parallax use of focused emitter and receivers, even off-the-shelf parts, could do a better job. All that is needed is a little custom circuit board.

The complexity in these components is comparing reflection intensity with and without the emitter on, in pulses, to filter background illumination. Perhaps that is the economy of these mass produced parts, containing all that logic. Still 8 pin PIC microcontrollers are pretty cheap. I would not know costs of programmable logic array parts, very popular in industry. Maybe analog components can do the trick. It is certainly easy to gate the signal with the emitter control pulses. The pulse-on and pulse-off signals could be accumulated over long exposure in separate capacitors for an analog comparator -- up to engineers.

[glnc222]

Maybe robots are just using the wrong parts. One sensor 22-70mm from Omron described as "convergent reflective" specifies
"Robust performance with little influence from sensing object color or material.
Detects objects with mirror reflective surfaces and transparent objects."
Omron B5W-LA01 data sheet https://www.omron.com/ecb/products/pdf/en-b5w_la01.pdf
Actually cheaper than the Neato Robotics Sharp sensor because lacks processing of the pulse signal, and just includes a power transistor for the emitter and output from the detector -- four wire instead of three. Then the cpu must do the processing. The Samsung Powerbot has four wire cliff sensors with a 496hz 40 per cent duty cycle from the system board and is defeated by some black materials even noted in the User Guide.

Omron does not explain how their sensor accomplishes its claims at least in the data sheet.

[edit] Omron described at
https://www.components.omron.com/produc ... la01/index

Looks like the parallax method, also given specification for particular distances, in or out of the spot.

[glnc222]

Roomba's new model i7+ September 2018
thread http://www.robotreviews.com/chat/viewto ... =1&t=20643
is described in an IEEE Spectrum review referenced as having solved the problem of black floors confusing cliff sensors. No detail was given but perhaps they have adopted a parallax instead of intensity type IR sensor discussed in this thread.