RoboHarness paper reports coding agent outperforming robot action models

Researchers report that Qwen3.8-Flash-Next outperforms state-of-the-art vision-language-action models on BEHAVIOR Challenge 2025 tasks when equipped with RoboHarness. The preprint presents a visual-geometric control panel that lets language-model agents understand and invoke robot tasks. The approach uses coding agents for embodied manipulation, contrasting with training models on robot-action data and video demonstrations. The researchers say they will release the implementation as open source.