RoboHarness: Simple robot framework outperforms VLA and world action models on BEHAVIOR tasks

The research team proposes RoboHarness, which gives LLM agents a visual-geometric control panel so they can directly understand and invoke embodied tasks. With this framework, Qwen3.8-Flash-Next surpassed the current state-of-the-art VLA models on BEHAVIOR Challenge 2025 tasks, and the implementation is open source. The authors believe coding agents have the potential to achieve generalist embodied manipulation without relying on large-scale training of vision-language-action models.