In the news
PAC-MAN: Perception-Aware CBF-RL for Whole-Body Safety in Humanoid Dodgeball
arXiv cs.AI · Published · 3 min read
In 30 seconds
- What happened
- Researchers developed PAC-MAN, a framework combining control-barrier functions with reinforcement learning to enable humanoid robots to dodge incoming objects using only onboard camera vision.
- Why it matters
- Roboticists building real-time safety systems for humanoids need practical perception-aware methods that work with limited onboard sensors rather than perfect state information.
- Watch out
- The approach succeeds on 95% of throws in testing but relies on semantic segmentation and specific camera setups; performance degrades significantly without accurate ball tracking or privileged information.
- eval
- benchmark
The patterns behind this
- Eval-Driven Development (Agent CI)
- MLCommons AI Safety Benchmark v1.0
- Reinforcement Learning from Human Feedback
Each one covers how the technique works, when it earns its cost, and where it breaks.
The Agent Architect
One pattern, one tradeoff, one production failure story. A short weekly briefing for people building agentic systems.
Weekly email, one-click unsubscribe. We only use your address to send the briefing.