Coupling control-barrier functions with onboard perception (rather than perfect state) is practical for real-world robot safety—the robot learns to dodge using only what its camera sees, not simulated perfect information.
PAC-MAN combines safety constraints with realistic robot perception to enable a humanoid robot to dodge balls. The system uses depth camera images and semantic segmentation to detect incoming balls, while control barriers ensure the robot's body parts stay safe. Testing shows the approach works nearly as well as an oracle with perfect information, and successfully deploys on a real robot.