Humanoid robots can walk, run, even do backflips. What they can't do reliably: pick up a screwdriver. The hand is now the gating constraint for the entire industry.
I love the article. It's nice to have some temporary references and magnitudes to check how this technology evolves along following months. Once read this article, I'm completely sure that the Chinese Humanoids Olympics are not longer about hands dexterity, but mainly about main body actuators and hardware.
Great piece. The hand-as-bottleneck framing is spot on. I think the deeper issue goes beyond hardware though. What makes contact-rich manipulation so hard is that it requires fine-grained prediction of physical interaction: what happens when fingers close around something deformable, when grip force needs to adjust in real time, when a tool slips mid-use. Current robotics foundation models are getting decent at gross manipulation, but the moment you need to predict what's happening at the contact surface between a fingertip and a screwdriver, you need a level of physical simulation nobody's cracked yet. Better hands alone won't get us there without the predictive models to drive them.
I love the article. It's nice to have some temporary references and magnitudes to check how this technology evolves along following months. Once read this article, I'm completely sure that the Chinese Humanoids Olympics are not longer about hands dexterity, but mainly about main body actuators and hardware.
Yes, It’s a serious technology barrier to the potential of humanoids. Here is my initial findings: https://www.the-waves.org/2020/08/04/robots-take-over-jobs-which-hands-to-prevent-them/
Great piece. The hand-as-bottleneck framing is spot on. I think the deeper issue goes beyond hardware though. What makes contact-rich manipulation so hard is that it requires fine-grained prediction of physical interaction: what happens when fingers close around something deformable, when grip force needs to adjust in real time, when a tool slips mid-use. Current robotics foundation models are getting decent at gross manipulation, but the moment you need to predict what's happening at the contact surface between a fingertip and a screwdriver, you need a level of physical simulation nobody's cracked yet. Better hands alone won't get us there without the predictive models to drive them.