HF Paper PyRUA-Lean Lifts Robot Task Success 14% While Cutting Tokens 65% With Python Agent Primitives

PyRUA-Lean, a new framework published on arXiv October 1, combines classical robot primitives with vision-language-action policies inside Python code cells that perform conditional checks and local retries. Across 700 simulated robotic tasks, success rate climbed from 63.1% to 71.7% while input tokens fell 65% and LLM calls dropped 49% on solved instances. The agent requests only the visual data and state feedback it explicitly needs for replanning, challenging the assumption that higher robot-agent success demands longer context windows.