saurabhprasun20/VLM-Oracle: Publication release for zenodo

VISOR: Initial public release This release introduces VISOR, a Vision-Language Model (VLM)-based test oracle for robotic system testing. VISOR evaluates robotic manipulation task executions from videos and supports both task correctness and task quality assessment. ([GitHub][1]) What's included Binary classification for task correctness, answering whether a robotic task execution is successful or failed. ([GitHub][1]) Multiclass classification for task quality, enabling quality-level assessment for successfully completed robotic tasks. ([GitHub][1]) Support for four robotic manipulation tasks: pick-up move-near put-in put-on ([GitHub][1]) Experiment execution through YAML-based configuration files using: experiment_runner.py binary and multiclass config files for different providers. ([GitHub][1]) Model and experiment support This release includes configuration support for: Gemini OpenAI ([GitHub][1]) The example configs also show evaluation setups across: openvla-7b pi0 spatialvla-4b ([GitHub][1]) Research scope VISOR is organized around three main research directions: evaluating how effective VLMs are as automated test oracles for robotic task execution, studying uncertainty in VLM assessments, analyzing how uncertainty correlates with oracle decisions for correctness and quality. ([GitHub][1]) Repository contents This release includes: experiment runners and helper scripts, binary and multiclass evaluation pipelines, result folders for Gemini and OpenAI runs, analysis tables, Jupyter notebooks in the RQ-Result folder for statistical analysis and final table generation. ([GitHub][1]) Setup Basic setup in this release: create a Python virtual environment, install dependencies with pip install -r requirements.txt, configure environment variables in .env, run experiments with: python experiment_runner.py ([GitHub][1]) Notes This is the initial release of the VISOR repository and is intended to provide the full experimental structure for running and analyzing VLM-based oracle evaluations in robotic manipulation settings.