Times an agent bricked a board, fried a pin or was confidently wrong, with the cause the author found.
Newest first, as we log them.
Filtered by
38 d ago
Wenbo Zhang and colleagues ran GPT-6 Astra as a zero-shot robot manipulation policy on the 42 simulated RoboDojo tasks. It ranked above all 40 public policies on that board. On real arms, the team halted testing after the model repeatedly issued unsafe actions. The paper says some incidents damaged hardware and nobody was hurt.
3 mo ago
Anthropic's red team had Claude, GPT, Gemini and other models drive robots, mostly in simulation plus a real Unitree Go2. The team reports that Opus 4.6 misjudged where a trash can was and the Go2 dragged it. Grok 4.1 Fast headed for a table reflected in a glass door, and no model completed an office loop on vision alone.
5 mo ago
Alice Plante had Claude write the code and choose actions for a SunFounder PiCar-X on a Raspberry Pi 5 and a FrodoBots Earth Rover. The Earth Rover's commands went through a cloud relay while Opus took about ten seconds to answer, and it drove into trees. A small binary on the rover cut command latency to under 10 ms.
