Researchers have introduced PAC-MAN, a framework that integrates control-barrier functions with reinforcement learning to address whole-body safety in humanoid robots. The work targets a dodgeball scenario, where the robot must react to an incoming ball while maintaining balance and avoiding collisions across its entire body.
A key feature of the framework is its perception-aware design. Instead of assuming the robot has full knowledge of the ball's state, the deployed policy works with what the onboard sensors actually see. This makes the safety guarantees more realistic for real-world deployment, where sensing is limited and noisy.
The paper is available on arXiv, though the abstract is truncated. Based on the provided text, the main contribution is coupling safety-critical control with learning-based policies under perception constraints, rather than proposing a new game-playing strategy. No other sources were included, so no comparisons or disagreements could be noted.