Twitxy Saved

Robots Digest 馃

Verified account This account is verified. Learn more

@robotsdigest

Follow @RobotsDigest for latest in Robotics, Humanoids, and Hardware + AI.
GET POSTS IN INBOX: robotsdigest.substack.com Joined August 2025
0 Following 6.5K Followers 3.4K Posts

Latest posts 5

Robots Digest 馃 Verified account @robotsdigest
What if a robot could practice inside a learned simulator, then perform better in a standard physics simulator?
PrimeSim does exactly that. It is a world model trained on 30,110 hours of demonstrations from 22 robot platforms, and GRPO training inside it raised OpenVLA-OFT's LIBERO success rate from 61.7% to 73.2% after 200 updates
0:08
1 replies 2 reposts 34 likes 1.4K views View on X
Robots Digest 馃 Verified account @robotsdigest
Video diffusion models have always struggled to deliver true real time interactivity.

Shuo Chen and the Roblox research team just solved this with PixVid. By lifting latent video diffusion models directly into pixel space, PixVid generates one frame at a time with zero VAE and zero temporal compression,

cutting median control to screen latency down to 34 milliseconds.
1:33
1 replies 3 reposts 34 likes 1.8K views View on X
Robots Digest 馃 Verified account @robotsdigest
Traditional 3D reconstruction requires hundreds of dense camera views and expensive capture rigs.

World Labs just obliterated this limitation with Atlas.

This multimodal world model turns a single photo or text prompt into a fully explorable 3D environment with pixel perfect camera control.
1:30
1 replies 5 reposts 12 likes 984 views View on X
Robots Digest 馃 Verified account @robotsdigest
What if an image generator was also a perception model?

Open Vision Banana fine tunes the open FLUX.2 Klein (4B and 9B) so it paints its predictions as RGB images.

One model handles semantic, instance, referring and reasoning segmentation, plus depth and surface normals. It also keeps its image editing ability.