GaP: A Graph-as-Policy Multi-Agent Self-Learning Harness For Variational Automation Tasks

The Core Claim

Robots can reliably handle real-world tasks with natural variation—like picking objects of different shapes or positions—by automatically generating and refining interpretable programs rather than training black-box neural policies. The authors show this works across 8 new benchmarks (4 simulated, 4 real-world), substantially outperforming existing baselines.

Why This Matters

Most robot learning falls into two camps. Hand-coded programs are reliable but brittle—they break when conditions change slightly. Neural network policies adapt to variation but fail unpredictably in ways engineers can't debug or fix, a critical problem for factories and warehouses. The authors tackle what they call "Variational Automation" tasks: real industrial work where object geometry, pose, or other conditions vary within a predictable range but traditional automation can't handle them. Current model-free learning struggles here because success requires persistent, reliable execution—one failure in a 1,000-task batch costs money.

How GaP Works

Graph-as-Policy generates directed computation graphs—think of them as flowcharts where each node is either perception (computer vision), planning (deciding what to do), or control (actually moving the robot). Rather than hand-coding these graphs, GaP builds them automatically from a library called MORSL (Modular Open Robot Skill Library) containing reusable perception algorithms, planners, and controllers.

The key innovation is parallel simulation rehearsal. GaP creates many candidate graphs with different structures and parameter settings, then runs them all simultaneously in an internal simulation against different versions of the same task (different object shapes, poses). It watches which combinations work best and iteratively refines the graph—changing which nodes connect to which, tuning thresholds, swapping out algorithms. This is closer to code synthesis than policy learning: the system is writing programs, not weights.

Because these graphs are made of recognizable algorithmic components (standard vision pipelines, motion planners, control laws), engineers can inspect why a graph fails and patch it. This interpretability is crucial for industrial deployment.

The Results

The authors tested GaP on 8 new benchmarks spanning tasks like bin picking, assembly, and object manipulation. The abstract doesn't specify exact success rates, but claims "significantly outperformed baselines." Real-world results (4 tasks) suggest the approach generalizes beyond simulation, which is the hard part in robotics.

The implicit comparison is against model-free policies trained with reinforcement learning or imitation learning—methods that work well in narrow, controlled settings but often need retraining or fail silently in new conditions. GaP trades off the sample efficiency of learned policies for the reliability and interpretability that industrial applications demand.

This work bridges a gap: it asks whether recent breakthroughs in code-generating AI systems (like large language models writing software) can solve robot automation differently. Rather than learning end-to-end, GaP assembles solutions from trusted building blocks and validates them in parallel, combining the best of both worlds—adaptability and reliability.