Researchers introduce Graph-as-Policy (GaP), a multi-agent coding harness designed to bridge the reliability gap in Variational Automation tasks by combining interpretable programming with model-free adaptability. GaP generates directed computation graphs containing perception, planning, and control nodes from a Modular Open Robot Skill Library (MORSL) and uses internal simulation to rehearse and refine these structures.
- The system iteratively refines graph structure and parameters in parallel to improve success rates and throughput.
- Evaluation on 8 new open VA task benchmarks, comprising 4 in-simulation and 4 in real-world scenarios, shows GaP significantly outperforms baselines.
This approach enables robots to execute persistent and reliable tasks in commercial and industrial applications despite variations in object geometry and pose.