Araav Nayak

Pointer is a review board of AI agents that reasons over data to debate why your robot failed, and lets the physics simulator decide who is right. When a robot fails, the visible symptom is rarely the cause, and the evidence is split across camera video, sensor telemetry and controller logs that engineers line up by hand. Pick a failed robot episode and ask, out loud, “Why did this run fail?” Three investigator agents, all running on Muse Spark, look into it at the same time, each from a different angle, and use SAM 3.1 to measure what they see. They commit to competing hypotheses, a skeptic agent cross-examines them, and every hypothesis still standing has its fix tested in MuJoCo. The one whose fix actually works wins. Pointer then explains the failure by voice while pointing at the evidence in every view at once: the object outlined in the video, a cursor on the telemetry, and the log line, linked by a pink thread. To show it isn’t scripted, “New mystery” simulates a failure with a random hidden cause and decoy warnings. The board diagnoses it blind, and the answer key opens only after it decides. Everything, including the agents, the simulation loop, the mystery generator and the interface, was built during this hackathon.