Research map / Vision-language and multimodal / Robot action and navigation
Saliency-Aware Quantized Imitation Learning for Efficient Robotic Control
Seongmin Park, Hyungmin Kim, Sangwoo Kim, Wonseok Jeon, Juyoung Yang, Byeongwook Jeon, Yoonseon Oh, Jungwook Choi
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 |
|---|---|---|---|
| replace SIS with vision-language keyframe detection for state weighting (ablation) | SQIL on OpenVLA · LIBERO (OpenVLA INT4 avg) | avg success rate (%): 72.6 ± 0.7 -> 73.2 ± 0.6 | Improved |
| add SIS-weighted QRD loss to QAT | OpenVLA INT4 (AWQ) · LIBERO-Spatial | success rate (%): 80.9 ± 0.7 -> 83.9 ± 0.5 | Improved |
| add SIS-weighted QRD loss to QAT | OpenVLA INT4 (AWQ) · LIBERO-Long | success rate (%): 47.3 ± 0.3 -> 49.2 ± 1.0 | Improved |
| add QAT to uniform QRD distillation (ablation) | QRD only · LIBERO-Spatial (INT4 AWQ init) | avg success rate (%): 69.5 -> 83.9 | Improved |
| add quantization-aware fine-tuning (QAT) over PTQ | OpenVLA INT4 (AWQ) · LIBERO-Spatial | success rate (%): 80.1 ± 0.5 -> 80.9 ± 0.7 | Improved |
| use calibrated AWQ instead of uncalibrated RTN | OpenVLA INT4 · LIBERO-Spatial (INT4 AWQ init) | avg success rate (%): 73.3 -> 80.1 | Improved |
| quantize ViT to INT4 only (ablation) | OpenVLA PTQ · LIBERO | success rate (%): 74.0% -> 73.4% | No clear effect |
| quantize LLM to INT4 only (ablation) | OpenVLA PTQ · LIBERO | success rate (%): 74.0% -> 71.3% | Hurt |
| use raw SIS values as loss weights instead of threshold weight alpha_t (ablation) | SQIL · LIBERO | avg success rate (%): 73.2 ± 0.6 -> 71.8 ± 1.0 | Hurt |
| change saliency noise distribution from Gaussian to Uniform (ablation) | SQIL · LIBERO | avg success rate (%): 73.2 ± 0.6 -> 73.0 ± 0.4 | No clear effect |
| compute SIS every 5th frame instead of every 4th (ablation) | SQIL on OpenVLA · LIBERO-Long | success rate (%): 49.2% -> 48.2% | Hurt |
| increase saliency weight beta from 2 to 5 (ablation) | SQIL on OpenVLA · LIBERO-Long | success rate (%): 49.2% -> 47.8% | Hurt |
Closest papers by meaning
- AutoQVLA: Not All Channels Are Equal in Vision-Language-Action Model's Quantization (ICLR 2026 · similarity 0.87)
- Quantization-Free Autoregressive Action Transformer (NeurIPS 2025 · similarity 0.87)
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.86)
- ActQuant: Sub-4-bit Action-Guided Quantization for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.86)
- Predictive Inverse Dynamics Models are Scalable Learners for Robotic Manipulation (ICLR 2025 · similarity 0.86)
- Unified Reinforcement and Imitation Learning for Vision-Language Models (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.
