RoboRACE Playground

Five cameras watch a robot arm; a vision-language model (Cosmos-Reason2-2B, run on a CPU) scores each view for “did the task succeed?”. Choose which cameras lie and see what each fusion rule concludes on 110 real cached episodes. The front camera is the receiver: the robot's own camera, which it trusts. All numbers are precomputed by the benchmark code in the RoboRACE dataset; nothing is simulated in your browser.

Robot
Attack
Lying cameras

Accuracy over all 110 evaluation episodes (thin line = 50%, coin flip)

Attacks: invert – liars report 1 − p (a consistent liar still carries information); collude – all liars send one identical inverted report; noise – random reports; stuck – always “success”. Methods are label-free: none sees the ground truth. Each camera's vote uses its own threshold (a quantile of its unlabelled history), so a low raw score can still count as a “yes” vote. Full study, code and videos: showcase · RACE-Net model.