Building robots since 1995; building robots with Codex since 2024 OpenAI Codex Ambassador - Northwest Arkansas
USA
Joined January 2025
- Tweets2.7K
- Following960
- Followers5.5K
- Likes13.7K
Pinned Tweet
gpt-6 astra is also handling all of the RL training/motion policy for the robot it has designed 😎
current progress with some basic obstacles/drops; i gave it a $25 budget for GPU rentals on vast.ai and it currently has ~$18 remaining
👀 @OpenAIDevs
instead of writing all those useless tests, why not have your working agent use gpt-6.1-terra as a subagent for end-to-end testing
it's lightning fast with computer use 😎
omg i'm sooooo over astra at this point; it's physically painful how slow it has gotten
6.1-sol is faster, but it hasn't proven to be capable/reliable for my workflows
what to do what to do
interesting that they're running the gating network and locomotion policy on a mac mini rather than something like a jetson
Robotic manipulation is moving toward adaptive, real-time physical interaction. Building on previous work where a Spot quadruped used whole-body control to manipulate a 15 kg tire, this system now operates autonomously outdoors.
By replacing external motion capture with onboard depth-camera perception and computing locally, Spot can perceive and manipulate objects in unmapped environments. Learn more about how this type of visual policy learning allows the robot to dynamically sequence and adapt behaviors like rolling, uprighting, dragging, and recovering from interference on the fly: rai-inst.com/resources/blog/…
ok ngl was initially disappointed that a new terra model was dropped rather than a frontier model
but...i decided to give it a go as my '/side' model and it's pretty good for that
very very fast on '/fast' mode which removes the frustration when needing to correct/guide/steer
quiet on X ≠ lack of progress; quite the opposite actually
i've somehow managed to completely burn weekly limits between the daily resets on the legacy 20x plan(s)
ask me anything lol
can't wait to upgrade my current robot controller (ipad mini 7) to this one 😍
New iPad mini was just released!
Looks like new Acoustic Display, OLED + nano texture, A20 Pro, and IP67 water resistance
Seems like decent upgrades apple.com/newsroom/2026/10/a…
rolie making stepping progress
seems like astra is spending a lot of time on the controller;
need to audit the whole process and likely guide it's sequencing
crazy that clip was only a year ago 🤯
now we've got codex designing a robot end-to-end; the engineering->cad design->motion policy
i'd say it's doing pretty well; here is a clip of the current motion policy
moving pretty good for a 20kg clanker right??
codex being better at predicting my next prompt than i am.....really shows me where the bottleneck lives
at this point i am consistently burning thru the codex weekly limit on 2x $200 plans
per day
anyone else or am i the oddball
hardware ppl: i need a STEVAL-VL53L9 board ASAP!
willing to pay double retail if anyone has an extra that they would sell
will be going on the chin of this robot
pls share with anyone that may be able to help!
ummmm idk abt this one
at this point, if your frontier 'sota' model can't do 3d, and 3d is the future (cad, ar/vr, robotics)....then maybe your sights aren't aimed in the right direction
The smartest code model in the world would not be good at 3d stuff unless it was trained specifically for it.
A lot of the labs recently got a big pile of 3d training data and an RL environment to improve 3d skills.
Take a look at the jump in 3d from Fable 5.1 to Opus 5.5. It's at least 4x better at 3d despite being a 5% better model at best.
Clearly Meta didn't go hard on 3d for this release. Doesn't mean the model is bad.
3dmaxxing is a different type of benchmaxxing, don't overthink it.
I'll take some fault here by sharing all the Fishslop demos. The 3d stuff is fun and cool as hell, but it's not a good representation of a model's real world code capability.