Re: [HBRobotics] vision

55 views
Skip to first unread message

Chris Albertson

unread,
Aug 20, 2026, 10:44:08 PMAug 20
to hbrob...@googlegroups.com


On Aug 20, 2026, at 2:54 PM, A J <aj48...@gmail.com> wrote:

Hey Folks,

The vision part of Bots is really exciting. It is amazing how nature has evolved over time.

Birds and snakes can see in a fourth color, like UV or IR. But mammals can compress

most of the information before it hits the optic nerve. I wonder if Bot cameras could do

the same thing. 


Yes, you can buy them. There are quite a few, but maybe the best known is “RealSense” sold by Intel.      There are others that are more generic and will run your own models.    Some of these do have an invisible to humans fourth color band, but that would be IR, not UV.     




Brian Higgins

unread,
Aug 21, 2026, 6:10:55 AMAug 21
to hbrob...@googlegroups.com
Yes, vision is amazing so is sound in humans
Brian Higgins
VA Researcher 
Guide Robot
Loomo Ride-on Robot 
Smart White Cane
40 year quest - sensor aided bike/scooter
Autonomous Vehicle for blind mobility “Laser enhanced Echolocation” & ClearPath Navigation 


Sent from my iPhone

On Aug 20, 2026, at 10:44 PM, Chris Albertson <alberts...@gmail.com> wrote:



--
You received this message because you are subscribed to the Google Groups "HomeBrew Robotics Club" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hbrobotics+...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/8F9CA42D-969F-4BBA-932C-324D3E9421C2%40gmail.com.

Dwell Pulse

unread,
Aug 21, 2026, 8:47:38 AMAug 21
to HomeBrew Robotics Club
hi

Chris Albertson

unread,
Aug 21, 2026, 12:04:40 PMAug 21
to hbrob...@googlegroups.com
Looks like I posted two identical links.      I wanted to send one to the RealSense camera and one to the OAK-D Camera.

OAK-D might be closer to what the OP was talking about.   I actually own this camera.   It really can mimic early vision processing if that is what is needed but can go further too in that it can do stereo vision and object detection in-camera.      It is really just a little box with a computer and three cameras.    The data from the three cameras can be combined inside the box.

I remember when the modern vision field was invented when David Marr published a book.  I think the title was just the word “Vision”.   It was foundational in the field.     It was also punished while I was a computer science student interested in and studying AI in the early 1980s.  Marr covers the whole process from neurology to psychology.   

I remember that I wrote off some of his ideas as impractical or impossible.   The field before Marr was about classical transforms and hand-coded pattern matching and nothing worked well. In the 80s, neural networks were an impractical sideline to “real AI”.   That changed in the 2000s after the video game era brought on GPU cards that could do linear algebra calculations billions of times per second.   To day, that GPU and the network can fit inside the OAK-D camera







-- 
You received this message because you are subscribed to the Google Groups "HomeBrew Robotics Club" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hbrobotics+...@googlegroups.com.

Stephen Williams

unread,
Aug 22, 2026, 2:02:36 PMAug 22
to hbrob...@googlegroups.com, A J

In the end, we'll probably just have a number of cheap cameras and a modest compute module at the expensive smart phone level.  While developing, having a range of smart sensors is probably desirable even if somewhat redundant.


The Oak-D is nice.  I ordered more than one of the LooperRobotics Insight 9 with built-in V-SLAM, 188deg FOV.

https://www.kickstarter.com/projects/looper-robotics/looperrobotics-insight-9-an-ai-native-robotics-camera/posts/4741561

I was considering the Pono P60-70M 3D Lidar sensor, but beyond the Insight 9's capabilities, I could use RGBD (image + depth).  The current nice lidars are about $700-800.  A lidar that provides RGBD is about $3000 now, and isn't that useful as the lidar data is always sparse compared to camera-style imagers.

They also have ultrasonic and laser distance sensors that could be useful:

https://www.positive-inno.com/


Because I need RGBD, the Pono guys pointed me to the NYX660/NYX650 for RGBD: Impressive, although I don't see a FOV listed.

https://www.amazon.com/ZICZNT-NYX650-Industrial-Waterproof-Dustproof/dp/B0DZSJ3FHL?th=1


Ouster has the RGBD lidar.  StereoLabs Zed cameras have been popular stereo-capturing cameras for several years.  They are now one company.  And they have a compute module with a Jetson Orin.  Nice combination, but expensive.

https://ouster.com/ouster-x-stereolabs

Their example use has 4 Zed X cameras, a lidar, and the ZED box to crunch all of that.  They include GMSL2-to-Ethernet processing which is nice.

https://www.youtube.com/watch?v=Gf1FyvUAdpc


These are the needs I'm trying to solve:

Where is the robot and the robot's end effectors (head, hands, feet, tail, tentacles) exactly in space relative to the room or landmarks?

What is the shape of the space the robot is in?

Where is everything around the robot

How can the robot / avatar operator be sure it is not running over something or someone or contacting anything it doesn't intend to?

Need high-quality feed of: 

  • high resolution video with human-like visual field, 
  • depth of that field if possible, 
  • stereo or enough information to generate a stereo view, and 
  • as wide field of video capture as possible beyond the range of the main visual field: 180 or 360 degrees ideally.

Need to be able to capture & compress video at a hardware level with HEVC, ideally stereo with MV-HEVC encoding.

https://www.youtube.com/watch?v=kHyY6JWCqYM

https://developer.nvidia.com/blog/enabling-stereoscopic-and-3d-views-using-mv-hevc-in-nvidia-video-codec-sdk-13-0/


Ideally just one sensor, at least in the forward direction, but likely 2+:

V-SLAM in a self-contained sensor is compelling, power saving.

If the V-SLAM device cannot run enough ML, might want another camera+ML processor like the Oak-D.  Or just use a Jetson board with enough power for everything.

Stereo capture cameras are probably the best way to get high-quality stereo video for teleoperation.  Until   Good RGBD might be usable, could be better, especially for scene understanding and another layer of spatial presence grid tracking for safety.  There is a latency minimizing reason this could be useful.

Lidar is a safe but expensive & limited range & resolution option, but provides 360 degree scanning.  Probably something to use early on for safety & training, then eliminate unless inexpensive at that point.


Stephen


On 8/20/26 2:54 PM, A J wrote:
Hey Folks,

The vision part of Bots is really exciting. It is amazing how nature has evolved over time.

Birds and snakes can see in a fourth color, like UV or IR. But mammals can compress

most of the information before it hits the optic nerve. I wonder if Bot cameras could do

the same thing. 


[source search engine]
Mammals compress visual data in the retina before sending signals to the optic nerve. This process reduces data volume by over 90%. Bipolar, amacrine, and retinal ganglion cells remove redundant background light levels, extract edges, track motion, and stream selective details through parallel pathways.
Key Types of Retinal Compression
  • Spatial Redundancy Reduction
    • Center-surround receptive fields compare light levels of adjacent points.
    • Uniform areas of light or dark are filtered out.
    • Only changes or boundaries at edges pass through.
  • Temporal Redundancy Reduction
    • Retinal circuits track changes over time rather than static images.
    • Constant, unmoving scenes fade out of the active signal.
    • Motion and sudden shifts trigger high-frequency firing.
  • Parallel Feature Extraction
    • Visual data splits into distinct functional channels.
    • Separate pathways process luminance, color contrast, and local motion.
    • Distinct retinal ganglion cells package these specific features independently.
  • Nonlinear Contrast Scaling
    • The system adjusts sensitivity based on overall ambient light.
    • Bright daylight and dim night inputs use logarithmic scaling.
    • This prevents saturation and preserves dynamic range within limited spike rates.


--
You received this message because you are subscribed to the Google Groups "HomeBrew Robotics Club" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hbrobotics+...@googlegroups.com.

Sergei Grichine

unread,
Aug 22, 2026, 6:08:51 PMAug 22
to hbrob...@googlegroups.com, A J
I have a pre-release OAK-D Lite (no IMU), with good power it works very well indoors and outdoors (RGBD + VSLAM). In case you want to experiment with OAK-D camera, here are my notes:


I tried to create an RGBD + AI/YOLO device using a spare Raspberry Pi 5 and a couple of cheap cameras. The obvious advantage is that all source code is under my control and can be tweaked to my liking.
The drawback is that the cameras need thorough calibration and it's a lot of work overall.


Visual SLAM is a lot of fun, here are my notes:


I hope this helps.

Best Regards,
-- Sergei


To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/6846b48e-1092-41b3-8b72-21c52a20cb60%40lig.net.

Asad Memon

unread,
Aug 22, 2026, 9:42:44 PMAug 22
to hbrob...@googlegroups.com, A J
I also ordered Looper Insight 9 and it’s nice in comparison with US alternatives. 

Might I shill my own SLAM camera I have been working on for 2 years now and significantly lower power req and price point: https://mightycamera.com

Stephen Williams

unread,
Aug 22, 2026, 10:47:41 PMAug 22
to hbrob...@googlegroups.com, Sergei Grichine, A J

Very nice.  How dense are the images from these?


Stephen

Stephen Williams

unread,
Aug 23, 2026, 12:33:21 AMAug 23
to hbrob...@googlegroups.com, Asad Memon, A J

Great.  Need to try that.  I remember an earlier mention.  Looks more interesting now.

This seems to solve the inside-out tracking that I was looking for.

I didn't find the FOV of your camera + lens combination.

Is some or all of your software open source or otherwise hackable?


The videos here work on Chrome, but only load the first segment on Firefox:

https://mightycamera.com/logs


This is an interesting solution for getting 4 synchronized cameras into a single MIPI-2 port:

https://www.uctronics.com/camera-modules/depth-cameras-for-rpi-jetson-nano-xaiver-nx/arducam-1mp-4-quadrascopic-camera-bundle-kit-for-raspberry-pi-nvidia-jetson-nano-xavier-nx-four-ov9281-global-shutter-monochrome-camera-modules-and-camarray-camera-hat.html


For a different purpose, no V-SLAM etc., I need a minimal camera (the OV9281 is good), but just enough of a microprocessor to receive the image & do simple processing.

Do you have recommendations for the most minimal, least expensive camera + microcontroller combination?

I can use the MightyCamera for now, but want to cut costs significantly for that application




Stephen

Reply all
Reply to author
Forward
0 new messages