Re: [HBRobotics] bots soon

[Dave Everett]

On Wed, 2 Sept 2026 at 03:05, A J <aj48...@gmail.com> wrote:

Hi Folks,

Tesla AIx Chip Specs (2030 Estimate)

Damn, based on past release claims, this means 2042.

Dave

[Chris Albertson]

It is all moot.  TOPS don’t matter.     Have you sever seem a robot whose behavior was constrained by a lack of compute power?

Today, any robot can have unlimited compute power if you can tether it to a big computer.    People do this all the time, and we still can’t see a robot replacing even low-skill human labor.

The bottleneck is the lack of good AI algorithms, not a lack of good enough hardware.   The better algorithms will come, but you know what?   I bet they are more efficient and use less computer power than current techniques.

[Karim Virani]

this is not moot. all my robots are compute constrained. my primary bot is connected to gemini live cloud and when it works nicely it's great. but the real world has network dropouts all the time - depending on where the robot is and how many redundant wireless connection technologies in use. what happens during longer duration dropouts? we have to fall back to local compute.

there are other kinds of compute (like obstacle avoidance) that you want to run as fast as possible with extremely low latency. machine vision workflows often want local compute.

that said, i'm not waiting on anything from tesla...

--
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/82F87E55-4738-496A-9D36-EDD976C97798%40gmail.com.

[Dave Everett]

I can recall back in the early 80s, people saying, when the 32-bit cpus come out, robots will have all the power they need. Seems nothing changes.

Dave

[Dave Everett]

What can your robot do Karim? I'm wondering how much could be done without connecting to other computers. I'm always trying to limit use of wifi etc.

Dave

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/CAKtnkixJa9Yyx2LEUBVFfz%2BOYB3L2zi1J%2BpL7noVNwpKJKvfSQ%40mail.gmail.com.

[Karim Virani]

ROS2 standard tooling can do quite a bit with traditional non or low ML algorithms. Mine is an outdoor focused sentry bot designed to roam my property and detect interlopers. It's a 4 wheel skid steer base (pioneer 3 AT) with a surround sonar array, 3D look-forward hemisphere lidar and a depth camera on the arm's end effector.

It uses Nav2 with robot_localization working from odometry+imu+rtk gps outdoors with very good accuracy when I have a good rtk fix. I have a different nav stack for indoor operations and am working on auto-detecting transitions between the environments. 

I also have an AI voice interface which uses the Gemini API - so it relies on-cloud resources. This interface is called By Your Command - and is my attempt to simultaneously support AI based goal generation simultaneously with general chat capability. But this only works when I have a low latency internet connection. My high quality rtk solution also requires internet - my usual source of correction data comes from rtk2go - though I could switch to a local base station over LORA with a bit more work.

I'm just using an Intel NUC box as my onboard compute.  This is a problem - I'm already near the limit with standard navigation and object avoidance. The minute I delve into more complex point cloud or depth frame processing (safe ground modelling, face recognition, etc) my update rate starts to suffer significantly.

If you're interested you can learn more about my sentry robot or find info about By Your Command on my projects page. Unfortunately this public documentation is a bit dated - I'm behind on making my latest work public.

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/CAG61pbcXnG8kX5eTCnmpSQx7zUh8_oewNfP2K02n_LxgbAhfhw%40mail.gmail.com.