The Missing Sense in Robotics: Tactile Intelligence for Physical AI| GelSight CEO Youssef Benmokhtar

Youssef Benmokhtar is the Chief Executive Officer of GelSight, the Waltham, Massachusetts company commercialising the vision-based tactile sensor invented at MIT. GelSight reads surface geometry down to a single micron, fine enough to resolve a fingerprint. In this episode he says plainly that industrial surface metrology is what pays the bills and robotics is a deliberate market-seeding bet that the company has chosen not to monetise. He explains that open-sourcing the DIGIT sensor was Meta's decision rather than GelSight's, and that the model is now under review ahead of the next-generation sensor: fully closed would be bad for the industry, fully open leaves value on the table. He argues against a single tactile foundation model, expecting separate low-resolution and high-resolution models instead, and concedes the low-resolution end to piezo and magnetic approaches on the record. He puts the tipping point for tactile intelligence at six to twelve months away, and points to fifteen years in the field against entrants that did not exist fifteen years ago as the commoditisation defence. He also covers the US Air Force SBIR grant to shrink GelSight Mini into a humanoid fingertip for aircraft and rotorcraft maintenance, the board additions of Rony Abovitz and Elliot Katzman, and the line he will not cross on weapons and law enforcement. Guest: Youssef Benmokhtar, Chief Executive Officer, GelSight (Waltham, Massachusetts) Host: Michelle Sun, Core Matter Full episode notes and transcript: https://corematter.substack.com/p/giving-physical-ai-the-sense-of-touch Core Matter, independent research on the Physical AI stack and supply chain: https://corematter.substack.com Chapters 00:00 GelSight: reading surface geometry to a single micron 01:14 Why an imaging executive took the CEO seat in 2021 03:37 Finding product-market fit outside robotics 04:51 Industrial surface metrology is what pays the bills 05:55 Reframing GelSight as a tactile intelligence platform 07:52 What makes a hardware product a platform 08:40 Where the durable advantage sits: optics, software, or data 09:42 Making the sensor easy so that people build on it 11:14 Rony Abovitz, and using "physical AI" before it was a term 12:32 Elliot Katzman: "you are not thinking about data enough" 13:55 Why open-sourcing DIGIT was Meta's decision 15:33 The tipping point, and what changed at the humanoid companies 17:51 Three years of CES: from humanoids doing flips to hands 18:58 Is the bottleneck hardware or tactile data at scale? 19:59 Why there is no ImageNet for touch 20:45 Why touch is harder to standardise than vision 21:34 The case against a single tactile foundation model 23:08 The US Air Force grant and a fingertip-sized sensor 25:20 Collapsing sensor costs and the commoditisation defence 27:52 Fifteen years against entrants that did not exist five years ago 29:23 When GelSight will build a custom sensor 30:44 The line on weapons and law enforcement 32:59 Nuclear plant maintenance as the clear humanoid case 34:36 "We are probably still six to twelve months away" 35:23 Sensor company, platform company, or data company in five years