PAC-MAN: Perception-Aware CBF-RL for Whole-Body Safety in Humanoid Dodgeball

PAC-MAN: Perception-Aware CBF-RL for Whole-Body Safety in Humanoid Dodgeball

|Authors: Lizhi Yang, Junheng Li, Aaron D. Ames

|Venue: arXiv preprint | Year: 2026

|arXiv: 2607.28623v1

Abstract

We present PAC-MAN, a perception-aware CBF-RL framework that couples control-barrier safety with deployment-realistic onboard sensing for whole-body humanoid dodgeball. The deployed policy sees the ball only as segmentation-masked depth from a head-mounted camera, while training-time CBF guidance represents clearance to every body link, and an adversarial motion prior regularizes the resulting evasive reflexes. We evaluate on a controlled any-link contact benchmark with seeded throws in two regimes: single throws and a deployment loop in which the robot walks back to its station and recovers between throws. On this benchmark, the policy comes within a few points of a privileged state oracle: a fixed onboard camera alone is adequate for evasion. We find that usable barrier structure depends on perceptual observability: Joint-CBF gives the best performance with accurate ball states, degrades under fixed-camera observations when used only as training guidance, and recovers with a ball-tracking gimbal or privileged runtime filter. We therefore deploy a lightweight Link-CBF policy zero-shot on the Unitree G1 in the real world, where it tolerates imperfect perception, succeeds on 95% of throws, and uses semantic segmentation to dodge different balls.

Key Contributions

• We present PAC-MAN, a perception-aware CBF-RL framework that couples control-barrier safety with …
• The deployed policy sees the ball only as segmentation-masked depth from a head-mounted camera, w…
• We evaluate on a controlled any-link contact benchmark with seeded throws in two regimes: single …
• On this benchmark, the policy comes within a few points of a privileged state oracle: a fixed onb…
• We find that usable barrier structure depends on perceptual observability: Joint-CBF gives the be…

Topics

• vision
• reinforcement-learning
• control
• benchmark

Code & Data

• GitHub repository: https://github.com/ccrpRepo/AMP_mjlab

BibTeX

@article{Yang2026_260728623v1,
  title     = {PAC-MAN: Perception-Aware CBF-RL for Whole-Body Safety in Humanoid Dodgeball},
  author    = {Lizhi Yang and Junheng Li and Aaron D. Ames},
  year      = {2026},
  eprint    = {2607.28623v1},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url       = {https://arxiv.org/abs/2607.28623v1}
}