PAC-MAN: Perception-Aware CBF-RL for Whole-Body Safety in Humanoid Dodgeball
PAC-MAN couples control-barrier-function safety with deployment-realistic onboard sensing for whole-body humanoid evasion. A policy trained with training-time CBF guidance and a head-mounted camera only achieves near-oracle dodge performance, and zero-shot deployment on the Unitree G1 succeeds on 95% of real-world throws.
Paper · arXiv:2607.28623Background
Safety guarantees for legged robots are typically derived from control barrier functions (CBFs) that assume accurate, privileged state information. In practice, however, a deployed humanoid perceives its environment through noisy onboard sensors — cameras mounted on the head, not an oracle feeding perfect ball trajectories. Bridging formal safety methods and real perception has remained an open problem, especially for whole-body tasks where every limb must avoid contact.
Core Innovation
PAC-MAN (Perception-Aware CBF-RL) couples barrier-based safety with perception-aware reinforcement learning. The deployed policy observes the world only through segmentation-masked depth from a head-mounted camera, while at training time a CBF module provides guidance based on clearance to every body link. An adversarial motion prior regularizes the resulting evasive reflexes, preventing the policy from collapsing onto unrealistic or unsafe movement patterns.
A key finding is that the usable barrier structure depends on perceptual observability: a joint-level CBF performs best with accurate ball states but degrades when used as training guidance under fixed-camera observations, recovering only with a ball-tracking gimbal or a privileged runtime filter. This motivated a lightweight link-level CBF policy that tolerates imperfect perception.
Results
On a controlled any-link contact benchmark with seeded throws, the perception-limited policy comes within a few points of a privileged state oracle, demonstrating that a fixed onboard camera alone is adequate for evasion. In the real world, the Link-CBF policy was deployed zero-shot on a Unitree G1 humanoid, where it succeeded on 95% of throws and used semantic segmentation to dodge different ball types.
Limitations
The evaluation focuses on a single humanoid platform and a dodgeball-style evasion task. The benchmark covers single throws and a walking-return deployment loop, but not the full range of dynamic disturbances a humanoid encounters in crowded or unstructured environments. The reliance on semantic segmentation for ball identification also assumes that objects of interest are visually distinguishable.
Industry Implications
Humanoid safety under realistic sensing is a central bottleneck for commercial deployment in warehouses, factories, and public spaces. PAC-MAN demonstrates that formal safety structure can be embedded at training time while keeping the deployed perception stack simple — a pragmatic path for manufacturers who want certified-style safety margins without expensive sensor suites. The any-link contact benchmark and the finding that lightweight policies survive imperfect perception are directly relevant to teams developing whole-body collision avoidance for bipedal and humanoid platforms.