There’s definitely a connection

5/
Check out the video in the first post of the thread to see one of the submissions in action.
The winning teams will be announced at the CVPR 2025 Perception for Industrial Robotics Automation Workshop.

Stay tuned!

4/
Some stats:
- 450+ teams registered
- Dozens of teams invested hundreds of hours
- Top 5 submissions were deployed on a real robot using Intrinsic Flowstate + ROS

We tried to pick parts based on participants' predicted poses - live!

3/
This is one of the toughest problems in computer vision, and it’s still far from “solved” - especially at the levels of accuracy and robustness required in real industrial settings.

2/
Participants were given multi-view, multi-modal images and tasked with training models that not only detect objects, but also predict their full 3D pose.
That means 3D position + rotation - exactly what a robot needs to grasp and manipulate objects accurately.

Ok, someone literally called their paper Foundation X: arxiv.org/pdf/2503.09860

Having your own pattern is a nice flex, so I’m happy for Bowen actually! :D

I only know Bowen Wen’s Foundation Pose and Stereo papers that follow that pattern. What else is there?

Find more details and apply here:
🔗 job-boards.greenhouse.io/intrinsicrob...
🔗 job-boards.greenhouse.io/intrinsicrob...

I’m wondering how long it’ll take until people start memorizing openings for all 960 starting options

And in million years you’ll get the answer 42

Here is the challenge website: bpc.opencv.org

The only reasonable explanation I can think of is that Nvidia’s moat is larger in training than in inference. If training becomes much more efficient, demand saturation becomes a risk. But even this explanation feels somewhat weak.