The Thinking Machine

Why Touch Is the Missing Piece in Robotics with Tao Yu

Vision got robots looking. Language got them reasoning. But the moment a robot has to actually do something with its hands such threading a cable, screw a nut on a bolt, or manipulate a deformable object it can't fully see, touch becomes the missing piece. And touch is one of the hardest unsolved problems in physical AI. In this episode of The Thinking Machine, I sit down with Tao Yu, Director of Dexterous AI Group at Analog Devices, to unpack why ADI is now building a humanoid hand platform with tactile sensing at its core. Here are just a few topics get into: - Why dexterous manipulation is the "crown jewel" of robotics, and why it's a hardware + sensor + data + AI problem all at once - What tactile sensing actually involves: pressure, vibration, temperature, and the multimodal fusion problem - Taxels (tactile pixels) and how human skin's ~1mm resolution sets the bar for fingertip sensors - Why collecting good manipulation data is so hard when the human operator can't feel what the robot feels - Real2Sim2Real with physics as the interface: how ADI thinks about modeling sensor non-idealities so policies trained in simulation actually transfer About Tao Yu: Tao Yu is the Director of Dexterous AI Group at Analog Devices, Inc. (ADI), where he leads research on multimodal tactile sensing for robotics and Physical AI. His work sits at the intersection of ADI's 60-year analog semiconductor heritage and the emerging demand for production-grade tactile sensors in dexterous manipulation and humanoid robotics. Tao holds a PhD from MIT and has published research spanning tactile sensing, wireless sensor platforms, and signal processing. At ADI, he and his team are pioneering industrial-grade tactile sensors that capture force, vibration, and temperature at sub-millimeter resolution, as well as Physical AI research on tactile-enabled dexterous manipulation. Tao on LinkedIn: linkedin.com/in/taoyumit/ [http://linkedin.com/in/taoyumit/] Thanks to Lightwheel for making this episode possible. Learn about how Lightwheel is making physical AI successful at https://lightwheel.ai [https://lightwheel.ai]

Physical AI and Robotics at NVIDIA GTC 2026 with Diana Wolf Torres

Diana Wolf Torres joins The Thinking Machine podcast to break down everything we saw at GTC 2026. Diana is the author of the DROIDS robotics newsletter and the Deep Learning with the Wolf AI newsletter, and we're both NVIDIA-invited creators who have been covering this space together since GTC 2025. We start with what stood out from Jensen's keynote — the shift to measuring tokens per watt, the Vera Rubin chip and why it might make your Blackwell obsolete before you install it, and NVIDIA's trillion-dollar hardware number. Then we get into OpenClaw, the DGX Spark, and why agentic AI is something you want to be experimenting with now. From there we hit the show floor and go company by company through the robotics that impressed us most — from robots filling real labor gaps to humanoids working in tandem. We also dig into robot safety, why it matters more than most people realize, and the role simulation plays in getting there. Timestamps: 00:00 - Intro 03:20 - Keynote Overview 09:29 - Build-a-Claw and OpenClaw 17:52 - Caterpillar's Autonomous Vehicles 21:48 - Unitree's H2 Robot 25:33 - Disney's Olaf Robot 35:13 - WORKR's industrial robots 41:55 - Psyonic's robotic hand and tactile sensing 43:30 - Humanoid and KinectIQ 49:08 - Noble Machines 52:15 - Generalist and GEN-1 57:11 - OpenMind 1:03:21 - Syncere - Lume 1:16:04 - Fauna Robotics and RIVR 1:20:19 - Agibot and Agibot World 2026 1:22:45 - Asimov About Diana Wolf Torres: Diana Wolf Torres is a Silicon Valley-based AI and robotics writer who publishes two Substack newsletters: Deep Learning with the Wolf, covering AI developments and ethics, and DROIDS!, focused on robotics news, founder interviews, and field reporting. She produces on-the-ground content from conferences like RoboBusiness and maker events across the Bay Area, translating complex technical developments for broad audiences. Diana on LinkedIn:  https://www.linkedin.com/in/diana-wolf-torres/ [https://www.linkedin.com/in/diana-wolf-torres/] DROIDS! Newsletter: droids.substack.com/ [http://droids.substack.com/] Deep Learning with the Wolf: https://dianawolftorres.substack.com/ [https://dianawolftorres.substack.com/] Website: www.droidsnewsletter.com/ [http://www.droidsnewsletter.com/] Thanks to Lightwheel for making this episode possible. Learn about how Lightwheel is making physical AI successful at https://lightwheel.ai [https://lightwheel.ai]

How Lightwheel is Building the Simulation Infrastructure of Physical AI with Steve Xie

Steve Xie spent years leading simulation at Cruise and NVIDIA before founding Lightwheel — and in that time he watched simulation go from a tool that was "great for showcasing to investors" to what he believes will become the core infrastructure layer for all of physical AI. In this episode, we sit down with Steve to break down Lightwheel's three-pillar framework for simulation infrastructure: World, Behavior, and Evaluation — and why getting all three right is what separates serious simulation from everything else. We also get into the physical measurement factory, the data scale that Lightwheel is hitting in 2025, and why RoboFinals may become the industry-standard benchmark for frontier robotics models. In this episode we discuss: - Why simulation started as a "toy" at Cruise and how Steve changed that. - The difference between a visually realistic asset and a physically accurate one. - Why Lightwheel operates one of the world's largest robotics arm factories. - How egocentric data and simulation data work together in the behavior layer. - The data pyramid: why real teleoperation is just the tip of the iceberg. - Why academic benchmarks are maxing out and what RoboFinals does differently. - How World, Behavior, and Eval form a flywheel — not just a stack. - The agentic core Steve sees sitting at the center of it all. - Why robotics data collection may eventually require a billion people. About Steve Xie: Steve Xie is the Co-Founder and CEO of Lightwheel. He brings over a decade of experience building simulation infrastructure across some of the most demanding environments in physical AI. Steve led the simulation department at Cruise during the early days of autonomous vehicles, then joined NVIDIA where he worked closely with the Omniverse team and developed his vision for simulation as next-generation physical AI infrastructure. He founded Lightwheel to build that infrastructure from the ground up. Follow Steve on LinkedIn: https://www.linkedin.com/in/stevexiecbs/ [https://www.linkedin.com/in/stevexiecbs/] About Lightwheel: Lightwheel is building the simulation infrastructure that physical AI needs to succeed — spanning world generation, behavior data, and evaluation. Their products include SimReady Assets, EgoSuite for egocentric data collection, and RoboFinals, an industrial-grade robotics evaluation platform co-developed with NVIDIA. SimReady Assets: https://simready.com/ [https://simready.com/] Learn more at: https://lightwheel.ai/ [https://lightwheel.ai/] Resources mentioned in this episode: LW-BenchHub: https://github.com/LightwheelAI/LW-BenchHub [https://github.com/LightwheelAI/LW-BenchHub] LeIsaac: https://github.com/LightwheelAI/leisaac [https://github.com/LightwheelAI/leisaac] IsaacLab-Arena: http://github.com/isaac-sim/IsaacLab-Arena [http://github.com/isaac-sim/IsaacLab-Arena] Thanks to Lightwheel for making this episode possible. Learn about how Lightwheel is making physical AI successful at https://lightwheel.ai [https://lightwheel.ai]

Why Robotics Is Harder Than It Looks with Chris Paxton

Robots can walk. They can dance. They can even do backflips. But can they reliably fold your laundry, make coffee, or recover from mistakes in your kitchen? In this episode, I sit down with robotics researcher Chris Paxton to talk about what’s actually hard about building intelligent robots. We explore: * Why robotics today is fundamentally different than it was 10 years ago * The rise of world models and robot imagination * Why contact and manipulation tasks are harder than navigation for robots * The compounding error problem in long-horizon tasks * Why robotics evaluation is still an unsolved challenge * How new data pipelines and egocentric data are accelerating progress If you’ve seen humanoids walking around conferences and wondered, “Are we really close?”, this episode brings clarity. Follow Chris on X: @chris_j_paxton [https://x.com/chris_j_paxton] Check out RoboPapers for deeper dives into robotics research: https://www.youtube.com/@RoboPapers [https://www.youtube.com/@RoboPapers]

Modeling the Real World with Tolga Kart

Tolga Kart spent seven years building massive 3D worlds for Call of Duty at Sledgehammer Games. Then he left gaming for Tesla Autopilot, led simulation at Parallel Domain, and now he's the CEO of Third Dimension AI, a company building neural simulators that reconstruct reality from sensor data. In this episode, we dig into SuperSim, Third Dimension's first product, which takes driving logs and reconstructs them into photorealistic 4D environments where robots can train and validate their behavior. The results look so real that Tolga has to convince people they're not just watching video. In this episode we discuss: - How SuperSim reconstructs real-world scenes in hours, not months - The difference between a "digital twin" and a "digital cousin" - Why procedural generation hit its limits for robotics simulation - The domain gap problem and why it's finally being solved - Generating synthetic edge cases: erratic drivers, collapsing bridges, kids running into the street - Why Gaussian Splatting is a good medium for robotics simulation - What's next for simulation for humanoids, drones, and beyond About Tolga Kart: Tolga Kart is the Co-Founder and CEO of Third Dimension AI. He brings over 2 decades of experience building cutting-edge technology in gaming, autonomy, and AI. Tolga began his career in gaming, shipping two Call of Duty titles at Activision Games before transitioning to autonomous vehicles. At Tesla, he built the Autopilot TPM team and rebuilt the simulation team, fully integrating it into Autopilot's development framework. Most recently, he led and scaled Parallel Domain's engineering organization in two years. Follow Tolga on LinkedIn: https://www.linkedin.com/in/tolgakart/ [https://www.linkedin.com/in/tolgakart/] Follow Tolga on X: https://x.com/tolgakart [https://x.com/tolgakart] About Third Dimensions AI: Third Dimension AI is a spatial generation company building the 3D worlds that will power tomorrow's embodied AI—from robots to autonomous vehicles—and enable new frontiers of creativity in gaming and entertainment. Third Dimension was founded in 2024 and backed by venture capital firms Felicis, Abstract, Soma Capital, MVP Ventures, and Solari Capital. To learn more, visit https://www.thirddimension.ai [https://www.thirddimension.ai] Thanks to Lightwheel for making this episode possible. Learn about how Lightwheel is making physical AI successful at: https://www.lightwheel.ai [https://www.lightwheel.ai]