Research map / Vision-language and multimodal / Robot action and navigation
Understanding the Impact of Geometric Foundation Models on Vision-Language-Action Models
Yurou Yang, Muyuan Lin, Roberto Martín-Martín, Labrie Martin, Shreekant Gayaka, Cheng-Hao Kuo, Luca Carlone
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 |
|---|---|---|---|
| fine-tune GR00T-N1.5 with VGGT tokens via Early Fusion, Late Fusion or Spatial Forcing (task finetuning only) | GR00T-N1.5 · RoboCasa (8 tasks, 3 cameras) | success rate (avg): 71.7 -> 69.7 (EF), 71.0 (LF), 68.3 (SF) | No clear effect |
| add Early Fusion VGGT tokens | GR00T-N1.5 · Unitree G1 real robot | overall success rate: 22.22 -> 27.78 (p=0.511) | No clear effect |
| add Early Fusion VGGT tokens | GR00T-N1.5 · Unitree G1 real robot (approach stage) | success rate: 57.78 -> 84.44 (p<0.001) | Improved |
| add mid-training on all RoboCasa tasks to Early Fusion (ablation) | Early Fusion · RoboCasa (8 tasks, 3 cameras) | success rate (avg): 69.7 -> 75.2 | Improved |
| mid-train baseline too and compare with mid-trained Early Fusion (ablation) | Early Fusion · RoboCasa (mid-trained baselines) | success rate (avg): 72.2 -> 75.2 (p=0.168) | Improved |
| mid-train Early Fusion (ablation) | Early Fusion · LIBERO | success rate (avg): 86.6 -> 87.7 (baseline 87.9) | No clear effect |
| use a single camera instead of three, comparing Early Fusion to baseline (ablation) | Early Fusion · RoboCasa (single camera) | success rate (avg): 17.2 -> 21.5 (p=0.030) | Improved |
| add learned attention gate to Early Fusion (ablation) | Early Fusion · RoboCasa | success rate: 5-27% -> 64-89% | Improved |
| make attention gate input-dependent (ablation) | Early Fusion · RoboCasa | success rate | No clear effect |
| add positional embeddings to Early Fusion (ablation) | Early Fusion · RoboCasa | success rate | Mixed |
| also finetune the LLM in Spatial Forcing (ablation) | Spatial Forcing · RoboCasa (8 tasks, 3 cameras) | success rate (avg): 68.3 -> 31.2 | Hurt |
| add VGGT tokens (Early or Late Fusion) to GR00T VLM for depth probing | GR00T-N1.5 · NYU Depth V2 depth probe | depth RMSE | Improved |
Closest papers by meaning
- What matters when building vision-language models? (NeurIPS 2024 · similarity 0.88)
- VGGDrive: Empowering Vision-Language Models with Cross-View Geometric Grounding for Autonomous Driving (CVPR 2026 · similarity 0.88)
- Let Language Constrain Geometry: Vision–Language Models as Semantic and Spatial Critics for 3D Generation (ICML 2026 · similarity 0.87)
- Unveiling the Mist over 3D Vision-Language Understanding: Object-centric Evaluation with Chain-of-Analysis (CVPR 2025 · similarity 0.87)
- VLM-3R: Vision-Language Models Augmented with Instruction-Aligned 3D Reconstruction (CVPR 2026 · similarity 0.87)
- 3D-VLA: A 3D Vision-Language-Action Generative World Model (ICML 2024 · similarity 0.87)
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.
