Learning by demonstration, not programming: Fangwei Li, BEAR Robotics · Closed Loop at GTC Berlin

A bottle or a crate can be in a different position or color every time, so you can't really program a robot for a task like this. Fangwei Li from BEAR Robotics shows KR1, their wheeled humanoid, which learns by demonstration instead: people show the task, and the robot learns from head camera video, wrist camera video and its joint data. He talks about KINO-2, BEAR's latest learning platform, about the 1,000 to 2,000 demonstration episodes it took to reach what you see here, why BEAR builds on wheels and not legs for now, and how much KR1 can carry: up to 100 kg on the base and 25 kg as a payload for the body. Guest Fangwei Li, Co-founder and CTO, BEAR Robotics: https://www.bearrobotics.ai Mentioned BEAR Robotics KR1: https://www.bearrobotics.ai/kinisi Links Full episode, Closed Loop EP 01 (NVIDIA GTC Berlin):    • Point your agents to the repository: NVIDI...   Website and new-episode emails: https://closedloop.fm Tihomir on LinkedIn:   / tihomir-opacic   Closed Loop on LinkedIn:   / closedlooppod   Instagram:   / closedlooppod   X: https://x.com/closedlooppod Closed Loop: How Robots Learn From People is a podcast of conversations with the people building physical AI. Hosted by Tihomir Opačić, CEO of Orange Hill. Recorded on the floor of NVIDIA GTC Berlin, 20–22 October 2026. Numbers in this interview are what the company told us on the floor (their own claims, not independently tested). Music: Quantized existence by Tihomir Opačić Wheels or legs: which would you choose for a humanoid that works in real buildings? Tell me in the comments. #PhysicalAI #Robotics #NVIDIAGTC #Humanoids #ClosedLoop