REAL

REAL (Robotic Embodied Agents Learning) is an agentic framework for open-world mobile manipulation by InternRobotics. It establishes sim-to-real-consistent environment APIs and integrates a simulated user for human-in-the-loop interaction. Introduces REAL-Bench spanning 241 tasks. Outperforms commercial closed-source VLMs with 56.9% success on interactive tasks. 78.3% end-to-end success rate on a physical dual-arm mobile robot. Accepted to ECCV 2026. Open-source code at github.com/InternRobotics/REAL.