Eidon AI shuts down: robotics data is "brutally hard"
Sam Padilla closed his robotics data startup after two years and open-sourced everything, and Ted Xiao says there are a single digit number of real buyers in the whole market.
Shown this week
Flex-pi predicts 3D geometry, not just video, and needs fewer demos
Ge Yan, Jesse Zhang and team trained a 6 billion parameter world action model that predicts 3D pointmaps and DINO features alongside color video, instead of only reconstructing images. RoboPapers says the result is a policy that is much more demonstration-efficient, generalizes well, and handles complex long-horizon tasks, covered in episode 106.
Chelsea Finn: reliability is the bottleneck, and robots have no RLHF moment yet
Chelsea Finn and Perry Dong published a blog post on what is missing in RL for frontier robotics models, framed around a question: what will be the RLHF moment for robotics. Finn says current models work okay when a person reviews the outputs, like drafting code, but reliability becomes the bottleneck as systems act with more autonomy and more trust.
Reliability is the one of the biggest open challenges in AI right now.
Teleop and other arguments
Eidon AI shuts down, and the field argues robot data has no buyers
Sam Padilla is shutting down Eidon AI after two-plus years collecting robotics data, saying the thesis was right but the business is brutally hard, and he is open-sourcing everything they built. Ted Xiao called it the most honest take on the market: a brief arbitrage window, saturated, with non-trivial entry costs and a single digit number of real buyers.
the market is saturated, entry costs are surprisingly non-trivial, and there are a single digit number of real buyers