Research map / Vision-language and multimodal / Multimodal LLMs
GeoMIND: A Benchmark for Spatial Understanding in Robotic Manipulation
Jinghe Wang, Xinrui Cao, Duo Wu, Chenghao Gu, Yong Zhong, Linjia Kang, Tianyi Xiong, Zhi Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces GeoMIND, a benchmark for evaluating spatial understanding (spatial reasoning and spatial memory) in robotic manipulation, per the abstract.
Problem
Robotic manipulation in open-world settings needs robust spatial understanding (identifying targets in complex layouts and adapting to scenes that evolve through interaction), but existing embodied benchmarks are not stated in the abstract beyond being insufficient.
Approach
A benchmark framing spatial understanding for manipulation as two capabilities: spatial reasoning and spatial memory. Details beyond the abstract are not available.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| specify target by name and visual cue (ablation) | pi0.5 · GeoMIND-Base | success rate (%): 11.3 -> 39.2 | Improved |
| add semantic augmentation to instructions (ablation) | six VLA policies · GeoMIND | success rate (%): 22.6 -> 14.2 (avg) | Hurt |
| add visual augmentation (ablation) | OpenVLA · GeoMIND | success rate (%): 25.0 -> 12.5 | Hurt |
| add visual augmentation (ablation) | pi0 · GeoMIND | success rate (%): 16.7 -> 16.7 | No clear effect |
| add visual augmentation (ablation) | InternVLA-M1 · GeoMIND | success rate (%): 37.5 -> 37.5 | No clear effect |
| add execution augmentation (ablation) | pi0 · GeoMIND | success rate (%): 16.7 -> 10.2 | Hurt |
| add execution augmentation (ablation) | InternVLA-M1 · GeoMIND | success rate (%): 37.5 -> 35.4 | Hurt |
| add spatial enhancement (InternVLA-M1 vs general policies) | VLA policies · GeoMIND-Base | success rate (%): 26.4 vs 20.7 (OpenVLA) | Mixed |
| use modular VLM grounding with scripted execution instead of end-to-end | COPA/MOKA · GeoMIND-Base | success rate (%): 4.0-16.9 | No clear effect |
| use Claude-Sonnet-4.6 backbone instead of Gemini-2.5-Flash | COPA · GeoMIND-Base | success rate (%): 4.6 -> 15.9 | Improved |
Experimental setup
Benchmarks: GeoMIND
Closest papers by meaning
- RoboSpatial: Teaching Spatial Understanding to 2D and 3D Vision-Language Models for Robotics (CVPR 2025 · similarity 0.90)
- SpatialClaw: Rethinking Action Interface for Agentic Spatial Reasoning (NeurIPS 2026 · similarity 0.90)
- Unfolding Spatial Cognition: Evaluating Multimodal Models on Visual Simulations (ICLR 2026 · similarity 0.89)
- From Where Things Are to What They Are For: Benchmarking Spatial-Functional Intelligence in Multimodal LLMs (CVPR 2026 · similarity 0.89)
- Knot So Simple: A Minimalistic Environment for Spatial Reasoning (NeurIPS 2025 · similarity 0.89)
- GravMAD: Grounded Spatial Value Maps Guided Action Diffusion for Generalized 3D Manipulation (ICLR 2025 · similarity 0.88)
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.
