Hesai Kosmo 3D reconstruction announced, captures spaces with LiDAR and 8K camera and generates 3DGS, meshes, depth maps in the cloud. Uses generative AI to fill unscanned areas and textures. Claims millimeter precision and up to 5× speed.
Key points
Capture: LiDAR plus 8K camera.
Output: 3DGS, meshes, depth maps generated simultaneously in cloud.
AI: Generative AI fills unscanned areas and textures.
Performance: Millimeter precision and up to 5× speed.
Sources
Generative value classifiers new Q-function type
Generative value classifiers are a new parametrization of Q-functions that don't require action conditioning. They may combine gracefully with large, pre-trained policies and learn/generalize better from a richer output signal.
Key points
Generative value classifiers are a new parametrization of Q-functions that don't require action conditioning.
They may combine gracefully with large, pre-trained policies and learn/generalize better from a richer output signal.
Sources
- Original post
- alexanderswerdlow.github.io - Linked in the post
HiP3R launches local 3D reconstruction system
HiP3R introduces a method to piece together local 3D reconstructions for consistent scenes across thousands of frames. A demo is available and the project, paper, and code are released.
Key points
Project: HiP3R website released
Paper: arXiv preprint 2610.31635 published
Code: GitHub repository HiP3R made public
Sources
- Original post
- kkakkkka.github.io - Linked in the post
- arxiv.org - Linked in the post
- github.com - Linked in the post
LLMSeg turns multimodal LLM into text-prompted image segmenter
LLMSeg turns an off-the-shelf multimodal LLM into a text-prompted image segmenter with no training, no segmentation head and no mask decoder. It segments images using LLMs and B-splines loops.
Key points
Turns an off-the-shelf multimodal LLM into a text-prompted image segmenter with no training, no segmentation head and no mask decoder.
Segments images with LLMs and B-splines loops.
Sources
- Original post
- abdullahamdi.com - Linked in the post
ArtHOI accepted to IEEE TPAMI, makes 4D data
ArtHOI accepted to IEEE TPAMI. It turns generated 2D videos into physically grounded 4D interaction data with articulated objects, human motion and contact. No 3D supervision needed.
Key points
Acceptance to IEEE TPAMI
Generates 4D interaction data from 2D videos with articulated objects, human motion & contact
No 3D supervision needed
Sources
- Original post
- arthoi.github.io - Linked in the post
Niulinx preparing for autonomous driving tests in Munich
Niulinx is preparing for autonomous driving tests in Munich. Munich is emerging as a robotaxi hub in Germany, following Minga, Waymo, Mobileye.
Key points
Niulinx is preparing for autonomous driving tests in Munich
Munich is emerging as a robotaxi hub in Germany, following Minga, Waymo, Mobileye
