Summary
Sentdex introduces Chonk, a DIY-style robot from Vulcan Robotics, and shows what he describes as his first successful run of picking an object off the floor and placing it into a bin, with a language model handling the planning, motion, and control. The run used DeepSeek V4.1 Flash for all of the decision-making.
He walks through why this platform suits real-world research: two grippers with their own cameras, arms that slide up and down the body to clear a bin, five cameras plus a 2D lidar, wheels that handle thresholds and clothing on the floor, and a mostly 3D-printed body that is easy to repair. He notes the limitation of five-degree-of-freedom arms and compares the practicality of this design with humanoids for household tasks.
The video then digs into how the system gets the model to reason about what it sees. Each camera view is different, so the setup uses overlays such as approach corridors and pickup zones to help the model understand how motion will play out. The setup was refined using a simulator and GLM 5.3 Flash, and Sentdex openly discusses bugs he found, including the bottom camera being overlooked, making this a candid look at LLM-driven robotics in practice.
📺 Source: sentdex · Published October 08, 2026
🏷️ Format: Hands On Build
