New work targets more capable robot agents and world models

AgenticRoboRL reports a LIBERO-PRO success increase from 18.1% to 71.0% for a 9B planner, while the Codex comparison also changes external-memory conditions. ALPS proposes a visual encoder combining physical state, language alignment, and pixel reconstruction, with reported LIBERO and planning results. A Google Cloud AI repost claims harness changes raised AutomationBench from 61.6% to 77.4% with the executor fixed; methods and independent validation remain limited.