Search papers, labs, and topics across Lattice.
This study introduces FailBench, a comprehensive benchmark designed to assess the reliability of Vision-Language Models (VLMs) in evaluating robot manipulation tasks, featuring 2,197 attempts across diverse real-world and simulated sources. The evaluation of 13 VLM-based detectors reveals that the best-performing model achieves only 0.77 mean balanced accuracy, with a notable decline in performance on tasks requiring contact-intensive assembly. Furthermore, the research identifies a systematic bias in VLMs towards predicting success under ambiguous evidence, while demonstrating that targeted input-level interventions can enhance detection accuracy by 2.4 percentage points without additional training.
VLMs struggle with robot task evaluations, achieving only 0.77 mean balanced accuracy, and even fine-tuned models often underperform compared to general-purpose counterparts.
Vision-Language Models (VLMs) are increasingly used to evaluate robot manipulation outcomes, but existing benchmarks offer limited evidence of cross-domain generalization. We introduce FailBench, a benchmark for robot failure detection comprising 2,197 manipulation attempts across 14 public sources (12 real-world, 2 simulated). In FailBench, 75% of failures occur naturally, and six real-world sources come from non-failure-detection datasets. Evaluating 13 VLM-based detectors, we find the best model achieves only 0.77 mean balanced accuracy. Notably, models fine-tuned for failure detection consistently underperform general-purpose VLMs and their own pretrained baselines. Performance depends heavily on required visual evidence: models approach saturation when outcomes depend on observable object motion, but degrade to near-chance (<0.60 balanced accuracy) on contact-intensive assembly tasks. Error analysis reveals a systematic bias toward predicting success under ambiguous evidence, which persists even with increased reasoning effort. Finally, we show that input-level intervention--spatially localizing and cropping outcome-relevant regions--improves the top detector by 2.4 percentage points without extra training.
