VLM Grounding Evaluation Results Grounding evaluation results for vision language models on robotics manipulation datasets. Part of the vlm info loss project studying how VLM connectors transform visual representations. Background Our embedding level analysis shows VLM connectors perform a compress then expand transformation: they sharpen dominant object representations while compressing secondary object category identity. All tested models converge to ~83% dominant category accuracy post connector despite wildly different pre connector baselines (69 84%). Object counting improves through the connector, ruling out simple information deletion — the connector reorganizes features non linearly for LLM consumption. Grounding evaluation tests the end to end pipeline: does spatial information that survives the connector also survive the full LLM decode into structured bounding box output? Protocol Two pass grounding on 8 robotics manipulation datasets (DROID, LIBERO, TACO Play, JACO Play, Berkeley Autolab UR5, MolmoAct, NYU Door Opening, UT Austin MUTEX) with 3 camera viewpoints (2 exterior, 1 wrist mounted): 1. Scene inventory — "List all objects visible in this image with bounding box…
We use cookies for essential functionality and analytics. You can accept or reject analytics cookies.Cookie policy