Here is the next test conversation between our fully autonomous AI powered robots! We are almost done with testing and soon our fully autonomous AI show will begin!
We learned a lot from our first autonomous video. In this next video we prompted the AI’s to talk about video games and provide richer supporting context.
Everything but the initial prompt was generated by our AIs leveraging a LLM using a RAG model installed locally on a Raspberry Pi. Key was both providing context (who they are and what they know) and a manufactured chat history. The robots stay within the context for the most part, but the context and the chat history never mentions Mario Bros.
There are also a number of changes to how they communicate. We found our speech to text (Vosk) might have been interfering with our text to voice (Piper). And, the time to convert speech to text was 1 second, so the robots would have long pauses while speaking to each other. It would take about 3 seconds from when a robot finished speaking to starting their response, which is fine in a casual setting (like our workshop), but made for awkward videos. To help solve this problem, we had the robots communicate via websockets instead of Vosk. This lets the robots decide on a response while the other robot is still speaking.
Thank you for watching and please like and follow us if you want more!
