Robotics Has Been Stuck for 70 Years — Deepak Pathak, Skild AI

Show a robotics demo from 1957 next to one from today, and most people can't tell which is which. Deepak Pathak, co-founder and CEO of Skild AI and a professor at Carnegie Mellon, argues robotics stalled because it was treated as a hardware problem rather than a problem of building a general brain. Collecting robot data by teleoperation at one example a minute, he calculates, would take the entire US population more than a century to reach GPT-3 scale. Skild's answer is an "omni-bodied" brain: one model for any robot and any task. It's pre-trained on scalable data like simulation and human video, post-trained on teleoperation, and improved by a deployment flywheel. Pathak shows it inserting AirPods with a simple gripper, learning from human video with under an hour of robot data, and cooking omelets on $4,000 arms with only a camera. He explains why climbing stairs is harder than a backflip. He also shows GPU assembly for NVIDIA's Houston factory, and a robot learning to walk on two legs in three tries after its other legs are disabled. Speaker info: X/Twitter: @pathak2206 (https://x.com/pathak2206) LinkedIn: https://www.linkedin.com/in/pathak22 Website: https://www.cs.cmu.edu/~dpathak Related links: Skild AI: https://www.skild.ai Timestamps: 0:00 AI's progress and the robotics hype 1:02 Robotics has been "almost here" for 70 years 1:27 A 1960s block-copying robot 2:17 Teleoperation in 1957 3:57 Why robotics is stuck: the missing general brain 4:27 Moravec's paradox 5:42 There's no internet of robot data 7:22 Skild's thesis: one brain, any robot, any task 8:22 Teaser: many robots, one brain 9:31 How to scale robot data: a career retrospective 12:01 The three properties of robot data that matter 13:21 Pre-training, post-training and the deployment flywheel 14:56 Why laundry folding is actually easy 16:17 The AirPods task 17:21 Learning from human videos 18:36 The robot's "dream": imagined training data 18:56 Cooking omelets on $4,000 arms 20:41 Truly end-to-end: camera in, motor power out 21:21 Why stairs are harder than backflips 24:50 Deployment: GPU assembly for NVIDIA 26:00 Package delivery to the front door 26:25 Safety: adapting when the robot breaks