Robotics
Posts, essays, courses and company news on robot learning and autonomy. Newest first.
Learn robot autonomy at Stanford, now open to everyone
Pavone and co-authors make Principles of Robot Autonomy (PoRA) free online: a textbook (arXiv, print soon from Cambridge University Press) built from his Stanford course series since 2017, which joins classical robotics and modern AI across the full stack of perception, state estimation, planning, control and learning. It comes with video lectures from Stanford AA274A (robot autonomy) and AA203 (optimal and learning-based control), Python and Jupyter exercises, instructor slides, and an "Ask the Book" AI study assistant (Stanford announcement).
Eka Robotics Raises $150M to Build Superhuman Intelligence for the Physical World
Pulkit Agrawal (MIT) and Tuomas Haarnoja (ex-DeepMind) raise a $150M Series A, led by a16z and F-Prime at a $1.45B valuation, for dexterous manipulation that is both general and fast. Instead of imitating human demos, their vision-force-action (VFA) model adds touch and force to sight and is trained by RL in simulation, AlphaZero-style; they claim about 1,500 picks an hour, around 2.5x a human.
How to Train Your Robot
Robotics has no internet-scale corpus, so Potter surveys how robot training data gets made: teleoperation, real-world RL, simulation, synthetic data, and human video or handheld UMI rigs. Today's robot datasets look roughly 1,000x smaller than LLM text corpora, and closing that gap could cost $10–100B, unless better architectures need less data.