Stanford’s PBD-AG enables long-horizon robots to build persistent world models
Stanford’s new PBD-AG framework lets service robots autonomously build and update 3D world models in dynamic environments
Researchers have introduced PBD-AG, a framework for long-horizon service robots that autonomously build persistent world models in unfamiliar environments. PBD-AG separates robot-verified stable fixtures from revisable dynamic object events, maintaining reliability-weighted object states with a geometric visibility gate to prevent false deletions under occlusion. Inspection viewpoints are chosen by a graph-conditioned policy that balances target coverage, travel cost, collision risk, and redundant observation. In simulation experiments, it achieved higher coarse-fixture F1 than capability-matched controls, along with stronger identity continuity and event recall, and a physical-robot demonstration showed integration with onboard sensing.
- PBD-AG autonomously bootstraps structural baselines from onboard exploration and maintains reliability-weighted object states across geometry, semantics, identity, existence, and support relations
- Simulation results show higher coarse-fixture F1 scores (vs. peers), improved identity continuity, and event recall, with a physical robot demo validating real-world feasibility
- Uses a geometric visibility gate and graph-conditioned inspection policy to reduce occlusion errors and optimize observation planning
Why It Matters
PBD-AG unlocks next-gen service robots to operate autonomously in dynamic environments, reducing manual recalibration and improving long-term reliability for tasks like warehouse logistics or elderly care.