AgenticRobotics automates robot-policy training with 0.001 false-promotion rate

In a new arXiv preprint, researcher Hang Yu introduces AgenticRobotics, a backend-independent control plane that brings the agentic coding paradigm—think Claude Code or Codex—to robot-policy improvement. The architecture mirrors how coding agents work: a main LLM controller orchestrates disposable workers through durable train–evaluate–improve transactions. But robotics has one critical difference: tools like trained policies, training pipelines, and data collection routinely fail. So AgenticRobotics measures and records tool quality at every call, and expires tools when the underlying artifact changes. Its core components include an immutable objective, controller-owned measurement, commit-keyed crash recovery, an evidence-graded skill library, and a standardized tool registry with a recorded call surface.

The system's headline claim is operational: operators can leave the loop because promotion is evidence-gated, state is recoverable, and capability quality derives from records—not because the loop picks better checkpoints than a human (on one tested lineage, it doesn't). The gates deliver measurable safety gains: false-promotion control of 0.001 per run versus 0.005–0.021 for shipped policy, anytime-valid decisions under optional stopping, zero lost or duplicate effects when processes are killed, and all six tested artifact-tampering classes caught by a signed verifier. This makes unattended robot-policy improvement practical for production environments, reducing the need for constant human oversight while maintaining strong reliability guarantees.