Vision-language models used as robotic critics, reward generators, and failure detectors need to understand how a manipulation unfolds, not only whether it succeeds. Yet fine-grained process supervision is rarely available in a unified form, and existing evaluations do not systematically diagnose these abilities. Robo2Process, introduced by researchers in a recent arXiv preprint, is a trace-grounded framework for building process-aware supervision from heterogeneous robotic execution data.

The framework aligns visual observations with source-native signals and annotations, extracts local process units (LPUs), and generates structured question-answer supervision through task-specific sampling, deterministic label construction, and ambiguity filtering. This yields RoboProcessBench, a diagnostic benchmark with 12 question families spanning static monitoring and dynamic reasoning, and ProcessData, a corpus of 57,892 QA pairs drawn from four data sources and 260 manipulation tasks, split into training and evaluation sets.

Evaluating a broad suite of vision-language models reveals uneven process understanding. Supervision from ProcessData-SFT improves Qwen2.5-VL-7B and InternVL-3-8B on local state, motion, progress, and primitive-aware cues. External evaluation on RoboRewardBench shows that Robo2Process also enables out-of-distribution process-aware reasoning, suggesting the framework offers a reusable pipeline, benchmark, and training substrate for this capability.