No new humanoid shipped this week. What surfaced instead were three separate bets on the same underlying problem: humanoid robots can walk reasonably well now, but they still cannot reliably sense what they are touching, grip an object without damaging or dropping it, or learn a new task without a library of human demonstrations. That gap, not top speed or payload, is what will separate humanoid platforms over the next few years.
Three bets on the same gap
Tesla: sensing skin
Tesla has filed a US patent application for a touch-sensitive skin that can be molded around robot fingers and palms, according to Robotics & Automation News. The pitch is a scalable way to detect contact and force across a hand, something most humanoid grippers still lack. A patent filing is not a shipped part: it shows Tesla thinks the problem is worth solving, nothing more. Buyers comparing platforms on our humanoid comparison page should treat this as a roadmap signal, not a current spec.
Skild AI: skip the demonstrations
Skild AI demonstrated a self-play training approach, using its S1 robotics foundation model, that let robots learn physical skills, including soccer, without human demonstrations, Robotics & Automation News reports. Demonstration data is expensive to collect and narrow in coverage, so a method that cuts reliance on it matters if it generalizes. The caveat: soccer is a controlled, repeatable task. Whether self-play transfers to the messier manipulation work a warehouse or home actually demands is still open.
Kyutai: a small, open model
Kyutai, working with Inria and Ecole des Ponts, released REKI, a 2.4 billion parameter vision-language-action model for robot manipulation, pretrained only on public data, MarkTechPost reports. That is notably small for this category, and public-data-only training contrasts with approaches built on proprietary demonstration fleets. A smaller, openly trained model lowers the barrier for integrators who do not want to license a closed stack.
Why hands, not height
These three items point at the same gap from different angles: sensing, skill acquisition and the control model itself. None of it concerns how fast a robot walks or how much it can carry. A humanoid that walks well but cannot feel what it is gripping, or cannot learn a task without a demonstration library, is not useful on a factory floor or in a home.
A former Nvidia robotics product marketing lead, now a VC, told TechCrunch Robotics that the industry is in its "Lego era," with basic building blocks, actuators and robotic hands, still being assembled rather than standardized. That matches what is in this week's news: the component layer is still being worked out, not the finished product. The Robot Report's own guide on choosing a BLDC motor for robotic arms and mobile robots makes a related point: picking the right actuator already depends on sensors, safety requirements and the platform type, before manipulation software enters the picture.
A new ADI whitepaper, covered by The Robot Report, examines the challenges and opportunities of moving physical AI out of labs and into production machinery: a reminder that a hand that works in a demo video still has to survive duty cycles on a real line. Buyers tracking cost alongside capability can follow list prices across platforms on our price tracker.
More competition, not less
Pudu Robotics, known mainly for delivery and cleaning robots used in restaurants and hotels, is expanding into semi-humanoid robots and quadrupeds, Robotics & Automation News reports. That is worth watching independent of the hands story: a company with an existing service-robot customer base and manufacturing scale entering the humanoid-adjacent category widens the comparison set buyers need to work through, beyond the usual established names on our best humanoid robots list.
The human-in-the-loop counterpoint
Not everyone frames this as a pure engineering race. A columnist at The Robot Report argues that humans remain central to developing humanoid robots and training their AI models, partly because it helps maintain control over the systems. That is a useful check on the self-play story above: Skild AI's demonstration cuts reliance on human data for one task, but it does not settle how much human oversight the training loop should keep more broadly. The two positions are not contradictory, but they pull in different directions.
What this means for buyers
None of this week's developments are products you can buy or deploy today. A patent filing, a training demonstration and a research model are all upstream of a commercial robot. But they point to where real differentiation is likely to land: dexterity, touch sensing, and how cheaply a platform can acquire new skills. When evaluating humanoid platforms, ask vendors about those directly, rather than taking walking speed or payload figures at face value. Our glossary covers the basics of vision-language-action models and related terms for anyone new to this space.