Research map / Vision-language and multimodal / Robot action and navigation
Uncovering Robot Vulnerabilities through Semantic Potential Fields
Som Sagar, Jiafei Duan, Sreevishakh Vasudevan, Yifan Zhou, Heni Ben Amor, Dieter Fox, Ransalu Senanayake
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use PPO instead of A2C or SAC as the vulnerability search policy (ablation) | RoboMD · Failure ranking, 3 tasks | avg accuracy score: 75.0 (A2C) / 52.2 (SAC) -> 80.7 | Improved |
| use RoboMD instead of VLMs for failure ranking | RoboMD · Failure ranking | accuracy: VLMs below 60% vs PPO 76.0-84.0% | Improved |
| add 5-shot in-context examples to GPT-4o (ablation) | GPT-4o · Failure ranking, Square | accuracy: 44.0% -> 48.6% | Improved |
| add contrastive loss with image+text embedding (ablation) | embedding · Embedding quality | MSE to identity: 0.8426 -> 0.1801 | Improved |
| add contrastive loss to image-only embedding (ablation) | embedding · Embedding quality | MSE to identity: 0.6495 -> 0.6179 | No clear effect |
| add text (CLIP) encoder to image embedding with BCE only (ablation) | embedding · Embedding quality | MSE to identity: 0.6495 -> 0.8426 | Hurt |
| fine-tune policy on RoboMD-selected failures | BC-lift policy · BC-lift | accuracy (%): 67.91 -> 92.83 | Improved |
| select failures with RoboMD instead of fine-tuning on all failures (ablation) | BC-lift policy · BC-lift | accuracy (%): 85.48 -> 92.83 with 1.3 GB vs 9.0 GB | Improved |
| compare model-based embeddings under lowering light intensity | embedding · Lighting sweep | cumulative embedding distance: 0.00 -> 0.67 monotonic | Improved |
Closest papers by meaning
- AHA: A Vision-Language-Model for Detecting and Reasoning Over Failures in Robotic Manipulation (ICLR 2025 · similarity 0.88)
- Failing Forward: Adaptive Failure-Informed Learning for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.87)
- RoboProcessBench: Benchmarking Process-Aware Understanding in Visual-Language Robotic Manipulation (NeurIPS 2026 · similarity 0.86)
- RePO-VLA: Recovery-Driven Policy Optimization for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.86)
- SAFE: Multitask Failure Detection for Vision-Language-Action Models (NeurIPS 2025 · similarity 0.86)
- SafeVLA: Towards Safety Alignment of Vision-Language-Action Model via Constrained Learning (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
