Why today's humanoids won't learn dexterity

Role of touch in dexterity

• Debate over whether fine touch is strictly necessary: some argue many tasks (grabbing a glass with gloves, teleoperated manipulators, “claw machines”) can be done mainly with vision and crude feedback.
• Others counter that humans still have substantial tactile/pressure feedback even through gloves and that many tasks (threading a nut, using a screwdriver, lighting a match, opening doors with tricky locks) really do depend on rich, fast touch cues.
• Several note touch may be especially crucial for learning a skill, even if once mastered it can be partly run “open loop” with expectations and prediction.

Learning, data, and simulation

• Some see no fundamental barrier: robotics can be trained with massive synthetic data and modern physics simulators; control networks can run at hundreds of Hz, far faster than human feedback loops.
• Others report that, in practice, high-fidelity sim‑to‑real for contact-rich manipulation is still very hard: modeling friction, deformation, brittleness, and variability of real objects is more difficult than just collecting real data.
• Discussion of “bitter lesson”: big models plus huge diverse data versus carefully engineered representations. Several argue robotics has not yet had its GPT‑scale investment or datasets, so it’s premature to claim limits.

Sensors, actuation, and hardware limits

• Agreement that human hands massively outperform current robot hands in sensor density, variety (pressure, vibration, stretch, temperature), robustness, and self-protection. Cheap, thin, durable, high‑resolution tactile “skin” is still missing.
• Some suggest using accelerometers and motor current as proxy force cues, but others point out this is still far from thousands of mechanoreceptors per hand.
• Muscles vs motors: muscles have superb torque, bandwidth, and paired antagonistic control; motors win on endurance and precision but struggle with impact resistance, torque density for small joints, and multi‑DOF joints.

Economics and scope of humanoid robots

• Strong theme: economics, not just capability, constrains progress. General humanoids must compete with specialized, already-profitable single‑task robots and redesigned “lights‑out” factories.
• Some argue a modest, non‑fully‑dexterous robot that can reliably pick boxes or stock shelves would already be hugely valuable; others note that even basic box handling in unstructured warehouses remains hard.

Environment redesign vs universal dexterous robot

• One camp expects environments, tools, and products to be standardized for robots (special handles, labeled boxes, robot‑friendly kitchens) rather than robots reaching human‑level dexterity.
• Critics reply that you can’t retrofit the entire messy legacy world (old buildings, infrastructure, repairs), so truly general workers must cope with human-designed artifacts—or remain confined to tightly controlled spaces.

Wheels, morphology, and locomotion

• Many agree wheels are cheaper, more robust, and easier to control than bipedal legs, but others emphasize the real world is full of stairs, curbs, and rough terrain where legs still shine.
• Broader point: insisting on strict human shape may be a mistake; more practical “animal-like” or hybrid forms (multiple legs, extra arms, wheeled‑leg hybrids) could win in real deployments.

Human vs artificial complexity

• Several comments stress how staggeringly capable biological systems are: dense multi‑modal sensing, self‑repair, plasticity, and the evolutionary “training” behind them.
• Some doubt we’ll ever fully match human general dexterity; others think it’s only a matter of scaling models, sensors, and compute, but acknowledge we’re many orders of magnitude away in data and investment.

Critiques of the article’s framing

• A few readers argue Brooks underplays the role of representation learning (e.g., in vision, where raw pixels are used) and overstates the need for hand‑engineered front-ends.
• One points out his description of speech recognition as still reliant on heavy handcrafted preprocessing is dated: modern systems often train much closer to raw waveforms.
• Others think he downplays the learning/control side (how robots will be trained on new tasks in new settings) in favor of focusing on sensors and mechanics.