Giving computers eye balls Recently multi camera AI linecalling and instant replay for pickleball
San Francisco, CA
Joined July 2009
- Tweets583
- Following2K
- Followers567
- Likes8.4K
The amount of training data Anthropic is going to get from pushing everyone to use cloud hosted artifacts for ML workflows will be wild. It's a nice time saver, but really annoying how they've defaulted to this.
An idea for the teams building in the dirt: @bedrockrobotics @prontoai @GravisRobotics @TeleoAI @terabaseenergy
Here's me conceptualizing how a lidar/stereo vision scan can yield a sort of "compression" of 3D space
by perceiving Aristotelian "forms" in world space to collapse the space into meaningful objects instead of raw data
this could then enable a git like diff of changes in space and enable low cost persistent memory of spaces
beautiful
Finally built a little prototype of something I've been wanting for years
Git for 3D spaces
Spatial diffing, a quick 3D visualization of change in space over time
For instance, automatically detecting the changed regions in these 2 splats / videos, noticing the addition of the chair the removal of a poster board.
Lots of inspection use cases
#robotics #slam #gaussiansplatting #vlm
How does an industrial robot know which part to grab from a bin?
Depth camera + iterative 3d model alignment example
I built a small system that takes one image from an industrial 3D camera plus the part's CAD file, and works out the exact position and orientation of every part in the pile. This can then feed into a robot arm for picking the part out of the bin.
In the clip, the overlay is the CAD model drawn at the pose the system computed. Where it lines up with the real part, the pose is right. It then marks the part it would go for first.
How it works: a rough alignment of the CAD model to the depth data, a fine alignment (ICP), then a check that renders the model at that pose and compares it with what the camera measured.
Data: XYZ-IBD bin-picking dataset (XYZ Robotics), xyz-ibd.github.io, CC BY-NC-SA 4.0. Pose estimation and rendering by me; clip shared under the same license.
How robots see in 3D
I fused LiDAR and stereo vision from an open data set from a campus robot from UT Austin
As the robot drives past people you can see the camera data / color from the scene of trees and people overlaid at the depth detected by stereo vision. Where the orange particles are shown is where the LiDAR detected the objects closer than the stereovision calculation.
Data: UT Campus Object Dataset (CODa), UT Austin AMRL, amrl.cs.utexas.edu/coda, CC BY-NC-SA 4.0. Stereo depth and rendering are mine; this clip is shared under the same licence.
Building a perception stack for a humanoid from scratch. Part 1: a simulator where it learns to stop for people.
I’m feeling that people are generally underestimating what will happen when the most effective people in the world at creating value end up 100x more efficient due to AI
Mannn just when I stacked up a bunch of fun tasks for the weekend, they cut it off. Good marketing for when it comes back online next week I hope...
The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees.
The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance.
Access to all other Claude models is not affected.
We apologize for this disruption to our customers. We believe this is a misunderstanding and are working to restore access as soon as possible.
Read our full statement: anthropic.com/news/fable-myt…
AI is good enough that the limiting factor is the ability of the human driver to keep up. Seems that nervous system ease and malleability will mediate one's ability to get the most out of AI as a fixed/stuck nervous system blocks the ability to imagine/abstract/envision imo.
There ought to be more efficient, request level fee structure for large model usage similar to priority+gas fees on Ethereum.
So if you are short on time and need a request processed faster you can specify a variable tip to get the request processed. @FAL @maharshii
