Research map / Vision-language and multimodal / Robot action and navigation
SaPaVe: Towards Active Perception and Manipulation in Vision-Language Action Models for Robotics
Mengzhen Liu, Enshen Zhou, Cheng Chi, Yi Han, Shanyu Rong, Liming Chen, Pengwei Wang, Zhongyuan Wang, Shanghang Zhang
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 |
|---|---|---|---|
| train specialized semantic camera policy (SaPaVe Stage 1) instead of using generalist VLM | semantic active perception · ActiveViewPose-200K | success rate avg: 72.7 -> 84.3 (vs Gemini-2.5-Pro) | Improved |
| use active head camera instead of fixed camera (ablation) | SaPaVe architecture · ActiveManip-Bench | avg success rate: 36.17 -> 74.83 | Improved |
| add wrist cameras to fixed head camera (ablation) | fixed-camera model · ActiveManip-Bench | avg success rate: 36.17 -> 52.33 | Improved |
| add wrist cameras to active head camera (ablation) | SaPaVe · ActiveManip-Bench | avg success rate: 74.83 -> 73.16 | No clear effect |
| use SaPaVe instead of fine-tuned GR00T-N1 / pi0 | real-world active manipulation · real-world 4 tasks | avg success rate: 53.75 / 45.00 -> 85.00 | Improved |
| add Stage 1 semantic camera pretraining (ablation) | SaPaVe · real-world 4 tasks | avg success rate: 53.75 -> 85.00 | Improved |
| add Stage 2 active manipulation finetuning (ablation) | SaPaVe · real-world 4 tasks | avg success rate: 66.25 -> 85.00 | Improved |
| use decoupled action heads instead of a unified decoder (ablation) | SaPaVe · real-world 4 tasks | avg success rate: 71.25 -> 85.00 | Improved |
| use camera adapter (LoRA) instead of full finetuning (ablation) | SaPaVe · real-world 4 tasks | avg success rate: 73.75 -> 85.00 | Improved |
| add Universal Spatial Knowledge Injection (ablation) | SaPaVe · real-world 4 tasks | avg success rate: 68.75 -> 85.00 | Improved |
| test on unseen objects, lighting and scenes | SaPaVe · real-world generalization | success rate: 90 -> 85-95 (Occluded P&P) | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Multi-SpatialMLLM — Multi-SpatialMLLM: Multi-Frame Spatial Understanding with Multi-Modal Large Language Models (CVPR 2026) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- ActiveVLA: Injecting Active Perception into Vision-Language-Action Models for Precise 3D Robotic Manipulation (CVPR 2026 · similarity 0.92)
- I-Perceive: A Foundation Model for Vision-Language Active Perception (NeurIPS 2026 · similarity 0.89)
- ACTIVE-o3 : Empowering MLLMs with Active Perception via Pure Reinforcement Learning (ICML 2026 · similarity 0.88)
- DreamVLA: A Vision-Language-Action Model Dreamed with Comprehensive World Knowledge (NeurIPS 2025 · similarity 0.88)
- Exploring the Limits of Vision-Language-Action Manipulation in Cross-task Generalization (NeurIPS 2025 · similarity 0.88)
- Object-Centric Prompt-Driven Vision-Language-Action Model for Robotic Manipulation (CVPR 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.
