{"type":"rich","version":"1.0","provider_name":"Transistor","provider_url":"https://transistor.fm","author_name":"The OPTIM Update","title":"Why Pixel-Based Robot AI Is an Expensive Detour | Chao Cao, Sancho","html":"<iframe width=\"100%\" height=\"180\" frameborder=\"no\" scrolling=\"no\" seamless src=\"https://share.transistor.fm/e/b7a37a9b\"></iframe>","width":"100%","height":180,"duration":1804,"description":"Advanced manufacturing has automated the individual process steps. CNC, 3D printing, photolithography - however complex the step, we can build a machine for it. The handoffs between those machines still run on people, and in cell therapy that means $300,000-a-year scientists spending their days loading and unloading equipment while contamination risk and throughput bottlenecks drive up the cost of every dose.\nChao Cao (Co-Founder & CEO, Sancho - CMU robotics PhD, autonomy lead for CMU's DARPA Subterranean Challenge entries, ex-Boston Dynamics AI Institute) makes the case that the dominant approach to robot intelligence, pixel-based VLA models trained on massive datasets, is an expensive detour. Sancho bets on 3D geometric world models and test-time reasoning instead, and runs the entire stack on a single onboard compute module.\nWe cover the lasagna analogy for factory workflows, the math on the most expensive labor doing the lowest-value work, what five nines of reliability does to the data question, Waymo vs. Tesla as a data-quality argument, three years of sending robots into tunnels and caves, the October-to-March sprint from incorporation to NVIDIA's GTC keynote, the oversubscribed seed round co-led by Fusion Fund and Catapult, and why regulated cleanrooms are an easier place to start than a grocery store.\nChapters:\n00:00 Intro01:10 Roadmap and why the company is called Sancho02:15 The lasagna problem: the gap between machines04:51 $300K scientists loading machines by hand06:04 Why fixed automation isn't the answer07:02 Bet #1: 3D geometry over pixels09:46 Bet #2: test-time reasoning over data scaling11:25 Running the whole stack on onboard compute13:11 Three years in the dark: DARPA SubT lessons16:15 October to GTC in five months18:35 The oversubscribed seed round19:01 Two founders, two autonomy worlds21:08 Why start in the most regulated environments23:48 Hot takes: Waymo vs. Tesla and wasted data25:58 What Chao wishes he knew before starting27:30 Hiring:...","thumbnail_url":"https://img.transistorcdn.com/MD1uw3xsUV8NIC9WNTbAWnAxAqctNuVEgSaMzKwMDRQ/rs:fill:0:0:1/w:400/h:400/q:60/mb:500000/aHR0cHM6Ly9pbWct/dXBsb2FkLXByb2R1/Y3Rpb24udHJhbnNp/c3Rvci5mbS9jYTFk/YTA0N2ZmOThmMmZl/ZGVjMjA3NWMwZDUw/YmRlMS5qcGVn.webp","thumbnail_width":300,"thumbnail_height":300}