Simon Kalouche

Bezos browsing through my thesis is all one can hope for from a masters thesis @JeffBezos

Egocentric human data collection is very scalable but the data we really need is the fine motor skills (mm-level finger position and tactile data) that really unlocks dexterous manipulation.

Deeply saddened by the passing of my friend, FedEx founder Fred Smith. Fred was a legend and pioneer of express delivery. He was an idol, inspiration and had a profound impact on me. He was passionate, brilliant, very kind and humble. He was a marine and true patriot. He…

Andy Zeng's trademark toss.

You don't need humanoids to do dexterous manipulation. Robots are limited by intelligence, not because they're not shaped like humans.

LLMs were trained on the corpus of internet text. There is no analogous large-scale dataset for robotics. Pi-0: robot model trained on ~1 year of robot data Qwen-2.5: LLM trained on ~100,000 years of text data That's 5 orders of magnitude difference.

Too bad this humanoid will lose its job to a faster, lower cost robot arm.

Writing has been on the wall. Self driving will surely win over human driving. Better consumer experience, far lower cost, more scalability and predictability. If I were @Google I’d be investing tens of billions of dollars per year to entrench myself before others caught up.…

Low cost dynamic robot test bed = truck bed

Robotics is going to be the biggest industry in the history of the planet. Hundreds of billions of robots of all shapes, sizes and descriptions -@pmarca

Shoutout to Waymo's founder @SebastianThrun In robotics, scaling from prototype to production takes a while.

Pretty amazing. Waymo surpasses Lyft and on track to pass Uber in next 12 months

Cool! Someone should scale this up in production.

Imitation learning has a data scarcity problem. Introducing EgoDex from Apple, the largest and most diverse dataset of dexterous human manipulation to date — 829 hours of egocentric video + paired 3D hand poses across 194 tasks. Now on arxiv: arxiv.org/abs/2505.11709 (1/4)

Unsupervised Visual affordance Learning!

This is a really fascinating thread about the differences between transformers and the neural network architectures we see in nature. Vision systems in nature are specialized: they have distinct, modular components which work in different ways, as opposed to being composed of…

We've definitely mastered dancing robots