🤖 YC has a Paper Club - every two weeks they review fresh articles on AI, and the latest edition is entirely about robotics.
YC's content is usually superficial: motivational lectures for founders, leaving you with two slogans in your head. Paper Club is different - it's a dinner for less than a hundred people in Mountain View, where the authors themselves present their articles, and the audience asks substantive questions. Recordings are posted on YouTube.
The host starts by saying that for the tenth year in a row, he's been promised "next year robotics will be solved": first AlphaGo, then MuJoCo, then Aloha, then diffusion policy, then VLA. 2026 was promised as the year of robots - we're halfway through, and so far it's the year of demos. He then lists four barriers that have yet to be overcome, the most underrated being sensorimotor skills. Humans have normal and tangential force, temperature, vibration, humidity, and friction coefficient assessment all over their skin: you find a charger in your backpack by touch, without looking. At best, a robot has one F/T coin at its fingertip and a camera on its wrist.
Next, five presentations. Memory in policy - without it, a robot washes dishes endlessly and burns a sandwich because it sees the world anew at every step. What should a robot even consider before acting - it turns out, selective reasoning works better than exhaustive, and valid reasoning doesn't mean useful. One policy for a 22 DoF hand, trained entirely in simulation and working zero-shot on unfamiliar tools. The economics of world action models: the model that performs best on benchmarks requires two GB200 robots costing $70,000 each.
The best presentation of the evening was from the founder of Rerun, and it wasn't about models. The thesis: "robotics application companies" are the new SaaS. One paying client, teleop instead of autonomy at the start, minimal proprietary hardware, replicating the client's environment in their office. And an observation worth everything else: in large labs, the most common takeaway from analyzing policy failures is not "we need a different architecture," but "we need to instruct those collecting data differently."
#robotics@rvnikita_blog #ai@rvnikita_blog #ycombinator@rvnikita_blog #research@rvnikita_blog