Research map / Vision-language and multimodal / Robot action and navigation
Spatially Guided Training for Vision-Language-Action Model
Jinhui Ye, Fangjing Wang, Ning Gao, Junqiu Yu, Zhu Yangkun, Bin Wang, Jinyu Zhang, Weiyang Jin, Yanwei Fu, Feng Zheng, Yilun Chen, Jiangmiao Pang
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 |
|---|---|---|---|
| add spatial grounding pretraining and spatial prompting | Vanilla VLA (Qwen2.5-VL-3B) · SimplerEnv Google Robot VM | success rate (%): 66.1 -> 84.6 | Improved |
| add spatial grounding pretraining and spatial prompting | Vanilla VLA · SimplerEnv WidowX VM | success rate (%): 54.7 -> 73.2 | Improved |
| replace vanilla co-training with spatially guided training (ablation) | VLA · SimplerEnv WidowX VM | success rate (%): 61.1 -> 73.2 | Improved |
| add spatial prompting during co-training (ablation) | VLA co-training · gradient alignment PSS | PSS: 0.25 -> 0.42 | Improved |
| add spatial pretraining with robotic grounding data (vs general grounding data) (ablation) | ST4VLA · Google Robot VM | success rate (%): 72.6 -> 84.3 | Improved |
| use general grounding data pretraining instead of none (ablation) | VLA · WidowX VM | success rate (%): 54.9 -> 65.2 | Improved |
| raise grounding vs action co-train ratio from 1:1 to 1:10 (ablation) | ST4VLA · Google Robot VM | success rate (%): 52.4 -> 80.7 | Improved |
| raise co-train ratio beyond 1:10 to 1:20 (ablation) | ST4VLA · Google Robot VM | success rate (%): 80.7 -> 78.3 | Hurt |
| scale spatial pretraining data from 0M to 3M pairs (ablation) | ST4VLA · SimplerEnv average | success rate (%): 61.4 -> 77.9 | Improved |
| use 0.5M pairs of spatial pretraining (ablation) | ST4VLA · SimplerEnv average | success rate (%): 61.4 -> 61.0 | No clear effect |
| replace unified prompt with random padding (ablation) | ST4VLA · SimplerEnv average | success rate (%): 77.9 -> 58.5 | Hurt |
| replace unified prompt with box/point/trace constrained prompt (ablation) | ST4VLA · SimplerEnv average | success rate (%): 77.9 -> 76.6 (box), 74.9 (point), 73.9 (trace) | Hurt |
Closest papers by meaning
- Beyond Data Scaling: Representation-Centric Pre-training for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.92)
- UP-VLA: A Unified Understanding and Prediction Model for Embodied Agent (ICML 2025 · similarity 0.92)
- FineVLA: Fine-Grained Instruction Alignment for Steerable Vision-Language-Action Policies (NeurIPS 2026 · similarity 0.92)
- PriorVLA: Prior-Preserving Adaptation for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)
- Seeing Across Views: Benchmarking Spatial Reasoning of Vision-Language Models in Robotic Scenes (ICLR 2026 · similarity 0.91)
- Knowledge Insulating Vision-Language-Action Models: Train Fast, Run Fast, Generalize Better (NeurIPS 2025 · similarity 0.91)
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.
