Why Today's Humanoids Won't Learn Dexterity (rodneybrooks.com)

A prominent critique argues that despite huge VC and corporate spending on humanoid robots, current approaches won’t yield human-like dexterity. The essay traces six decades of manipulation research (from MIT and Waseda to Boston Dynamics and Rethink), shows that deployed industrial systems still rely on simple parallel-jaw grippers and suction, and stresses that many human-like hands developed in labs haven’t translated to robust, general-use dexterity. Startups are betting on end-to-end learning from video or teleoperation demonstrations, but those datasets are sparse, low-fidelity, and biased to a few repeatable tasks—Benjie Holson’s “humanoid Olympics” example highlights dozens of everyday skills current systems can’t generalize to. Technically, the piece spotlights concrete failure modes: lack of wrist/force feedback and rich tactile sensing, limited independent finger control, medium precision (~1–3 cm), and teleoperation that can’t convey force or fine touch. Those constraints explain why learned controllers succeed on narrow scripted tasks but won’t discover the robust contact-rich strategies humans use. The implication for AI/ML is stark: scaling data and compute alone won’t substitute for better hardware sensing, compliant actuators, and control paradigms that fuse touch and force with learning. Expect useful humanoids in the future, but likely with different morphologies, specialized end-effectors, or fundamentally new sensing/control breakthroughs rather than today’s humanoid prototypes.