Research map / Vision-language and multimodal / Robot action and navigation
Spatial Memory for Out-of-Vision Manipulation in Vision-Language-Action
Pengteng Li, Weiyu Guo, He Zhang, Tiefu Cai, Xiao He, Yandong Guo, Hui Xiong
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 persistent spatial memory with construction, refinement and retrieval (full SOMA) over scan-then-reactive policy (ablation) | GR00T N1.5 on real-world OOV tasks · real-world OOV tasks (avg of 5) | success rate (%): 18.5 -> 28.3 | Improved |
| add multi-view head scanning to build the initial memory (ablation) | SOMA · real-world OOV tasks (avg) | success rate (%): 19.8 -> 24.1 | Improved |
| add dynamic memory refinement during manipulation (ablation) | SOMA with scan-built memory · real-world OOV tasks (avg) | success rate (%): 24.1 -> 28.3 | Improved |
| scanning alone with a reactive policy (no persistent memory) (ablation) | GR00T N1.5 · real-world OOV tasks (avg) | success rate (%): 18.5 vs 19.8 for No-Scan SOMA | No clear effect |
| use spatial memory instead of GR00T-N1.5 reactive control | OOV Task 5 · time-to-grasp | seconds: 36.5 -> 14.6 | Improved |
| add geometric cues to memory (ablation) | SOMA · RoboCasa GR1 (Full) | overall success (%): 45.1 -> 49.3 | Improved |
| add object semantics to memory (ablation) | SOMA · RoboCasa GR1 (Full) | overall success (%): 43.7 -> 49.3 | Improved |
| add dynamic update mechanism (ablation) | SOMA · RoboCasa GR1 (Full) | overall success (%): 41.5 -> 49.3 | Improved |
| use a fixed head camera instead of an active head camera (ablation) | SOMA · real-world OOV tasks | success rate | Hurt |
| add SOMA memory in fully observable simulation with more demos | GR00T N1.5 on RoboCasa GR1 · RoboCasa GR1 (Full data) | average success (%): SOMA 52.0 at 300 demos but 49.3 at Full; GR00T 44.3 -> 45.9 | Mixed |
| SOMA memory on SimplerEnv Open/Close Drawer | SOMA vs OpenVLA-OFT · SimplerEnv Visual Matching, Open/Close Drawer | success rate (%): 47.2 (OpenVLA-OFT) vs 31.5 (SOMA) | Hurt |
Closest papers by meaning
- SAM2Act: Integrating Visual Foundation Model with A Memory Architecture for Robotic Manipulation (ICML 2025 · similarity 0.89)
- Localizing, Structuring, and Rendering: Bridging 3D and 2D Vision-Language-Action Models for Robotic Manipulation (CVPR 2026 · similarity 0.88)
- Seeing Across Views: Benchmarking Spatial Reasoning of Vision-Language Models in Robotic Scenes (ICLR 2026 · similarity 0.88)
- OmniManip: Towards General Robotic Manipulation via Object-Centric Interaction Primitives as Spatial Constraints (CVPR 2025 · similarity 0.88)
- VLA-OS: Structuring and Dissecting Planning Representations and Paradigms in Vision-Language-Action Models (NeurIPS 2025 · similarity 0.88)
- ConsisVLA-4D: Advancing Spatiotemporal Consistency in Efficient 3D-Perception and 4D-Reasoning for Robotic Manipulation (CVPR 2026 · 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.
