https://teamknox.com/en/tech/robotics/robocar/physicalconnect/index.html 　TriClops’ main computer is a Raspberry Pi 4. Various AI agents can run on the board. The goal is to control the robot using natural language. This includes direct commands, multi-step commands, and commands executed by writing scripts on the fly. Target of Operation Control the actuators, sensors, and other components installed on TriClops. By making the various scripts and programs used in RVCS available to AI agents, you can control them through those functions. Hugo en-us https://teamknox.com/en/tech/robotics/robocar/physicalconnect/log/index.html Mon, 01 Jan 0001 00:00:00 +0000 https://teamknox.com/en/tech/robotics/robocar/physicalconnect/log/index.html Since this is a constant process of trial and error, I’ll jot down my observations as I go. 2026/10/13 Gemini now starts up faster than before. I observed an interesting phenomenon, so I’m making a note of it here. Currently, TriClops has a photo-taking feature. I used it to take a very close-up photo of the Pokémon Gengar, then asked the web-based Gemini and TriClops-like RasPi4 Gemini agents what each of them saw in the image. The Web-Gemini correctly recognized Gengar. On the other hand, the TriClops agent claimed to see a structure shown in TriClops. That element isn’t in the image at all. This is the first time I’ve encountered such a strong hallucination. Even when I urged it in a neutral tone, saying, “Take another look,” the result was the same. Losing my patience, I urged it in a much stronger tone, “I said take a good, hard look!” and it replied, “It’s Gengar.” When I try reflecting with the AI itself, it seems that consciously using stronger language helps steer my train of thought in the right direction. Alternatively, telling it to “Start from scratch!” also seems to work well. The inertia of my “prior knowledge”—or whatever you want to call it—seems to be greater than I expected…