How We Created a Thinking Robot?
summary
A team builds robots that run a local on-board LLM and combine it with fast control models, showing a prototype that solves the trolley problem and a wheeled $1,000 robot that actually exits a room. Visual‑action (VA) models output action tokens 30×/s versus LLMs with multi‑second delays, so manipulation policies are trained from 50–100 teleoperation episodes (supervised) and stitched together by a controller; an Robocrew open-source framework wires camera augmentation, sensors and tool models but generalization and 3–5s latency remain limits.
AI-written summary · about
chapters (21)
- 0:00Thinking robots intro
- 2:18LLMs to robotics
- 5:39AI agent for robots
- 7:34Humanoid autonomy demo
- 9:16Real-world robot apps
- 12:15Visual language action models
- 14:30Affordable non-humanoid robots
- 16:48AI navigation challenges
- 18:59AI-controlled robotic arms
- 21:17Robot generalization challenge
- 23:46Autonomous robot systems
- 26:14Robocrew framework
- 27:17Q&A military use
- 29:37Home robots timeline
- 31:26Training visual models
- 33:16Wheels vs legs design
- 34:19Robotics hardware basics
- 35:23Robot touch sensitivity
- 36:54Camera feedback with LLM
- 38:28Humanoid vs specialized robots
- 41:56Closing thoughts
related talks
16mReal-World Robotics as the Next Frontier for AI?
19mThe Physical Turing Test: Solving General Purpose Robotics
30mThe robot in my living room - insights from a journey through the possibilities...
34mLLM as a Robotic Brain: Cloud-Driven Robot Action Sequences Generated by Large...
26mGating Cognition: How Modern AI Architectures Minimize Runtime Reasoning
14mNarraGuide: an LLM-based Narrative Mobile Robot for Remote Place Exploration
10mMeet the Speakers
32mThe Road to General AI