From a $5 chip to an NVIDIA award: 5 robots running AI with zero internet

We went back into the DIY-robotics research behind last week's roundup and picked the five builds that stuck with us most — as a countdown, this time, with a companion video made with MegaBrain.

Watch on YouTube → — full sources and timestamps in the video description. Made with MegaBrain; sponsored by privateSLM.

The countdown

#5 — LlamaPi (zhoupingjay): a Raspberry Pi 5 robot arm that gestures from conversation, running Llama 3.2 3B at roughly 3 tokens/second. github.com/zhoupingjay/LlamaPi

#4 — Chatbot Zero (Jdaie): a Raspberry Pi Zero 2W hit 68°C and went sluggish running a small model; the same builder's Raspberry Pi 5 follow-up was dramatically faster — the honest limit, and the fix. hackaday.io — Chatbot Zero

#3 — SenseCAP Watcher (Vivek Verma): an $89 camera+mic device running a multimodal model over the local network, 2–4 second response time, no cloud vision API. Vivek Verma — SenseCAP Watcher assistant

#2 — GLaDOS (Dave Niewinski): an animatronic robot personality on an NVIDIA Jetson AGX Orin with a robotic arm and custom-trained voice — winner, NVIDIA's own AI Innovation Challenge. hackster.io — GLaDOS animatronic

#1 — esp32-llm (DaveBben): a 260,000-parameter model running directly on a $5 ESP32-S3 chip, about 19 tokens/second by the builder's own account. Not the most powerful thing on the list — the clearest sign of where the rest are headed. github.com/DaveBben/esp32-llm

Full source list, specs and video-production notes for all 25 builds we researched: the original roundup and our research brief.

Every one of these builders is working around the same idea: keep the model close to the data, not the other way around. That's what privateSLM does on your phone.

Discuss this on the forum → — which of these five would you rank #1?