better than the average, unpracticed human*.

the first home robots will already be better than humans at some things they do, much like LLMs, a jagged form of intelligence.

diffusion policy with relative actions, if not normalized per timestep, has the inductive bias of generating later actions first.

i was surprised to find out that the ceo of unitree consistently has bad takes.

RT @gazorp5: @DJiafei Don't you think Diffusion Policy was GPT-1?.

toyota lbm architecture is weird.

wozniak’s coffee test.

RT @gs_ai_: Today, We’re launching Genesis AI — a global physical AI lab and full-stack robotics company — to build generalist robots and u….

someone get to the bottom of this, free research paper.

also note that a linear probe not being able to reconstruct the original image doesn't mean the information isn't there, it could just be that it's not easily decodable.

here's a colab link to reproduce.

i wonder if it's related to the pooling vs [CLS] token? . also, please do not ask me how i discovered this.

how much does the last layer of a VLM retain the original image?. i trained a linear probe on the output features of several CLIP/SigLIP models on image reconstruction, and found that *only* with SigLIP, if you multiply the input by 10-100, pixels get reconstructed perfectly??

RT @moayedhajiali: Where are good old progressive diffusion models? 🤔. Breaking generation to multiple resolution scales is a great idea, b….

RT @moayedhajiali: Big news 📢📢.AV-Link has been accepted to ICCV 2025 🚀🚀.

a great research codebase is one that puts you one config change away from a cool experiment.

RT @Haoyu_Xiong_: Your bimanual manipulators might need a Robot Neck 🤖🦒. Introducing Vision in Action: Learning Active Perception from Huma….

robotics is a hardware problem

RT @Dazitu_616: 📢 Introducing DenseDPO: Fine-Grained Temporal Preference Optimization for Video Diffusion Models. Compared to vanilla DPO,….

bro doesn’t have a product but boasting about internal reorg.