Research map / Vision-language and multimodal / Robot action and navigation
Towards Long-Horizon Vision-Language-Action System: Reasoning, Acting and Memory
Daixun Li, Yusi Zhang, Mingxiang Cao, Donglai Liu, Weiying Xie, Tianlin Hui, Lunkai Lin, Zhiqiang Xie, Yunsong Li
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 cold-start curriculum SFT (ablation) | Qwen2-VL-7B reasoning layer · Coarse-grained long-horizon tasks | accuracy (%): 29.00 -> 62.00 | Improved |
| add cold-start curriculum SFT (ablation) | Qwen2-VL-7B reasoning layer · Zero-shot generalization | accuracy (%): 19.00 -> 43.00 | Improved |
| add iterative DPO iteration 1 (ablation) | reasoning layer after cold start · Zero-shot generalization | accuracy (%): 43.00 -> 55.00 | Improved |
| add iterative DPO iteration 3 after iteration 2 (ablation) | reasoning layer · Zero-shot generalization | accuracy (%): 64.00 -> 73.00 | Improved |
| add iterative DPO iteration 3 after iteration 2 (ablation) | reasoning layer · Coarse-grained long-horizon tasks | accuracy (%): 75.00 -> 87.00 | Improved |
| add iterative DPO iteration 3 after iteration 2 (ablation) | reasoning layer · Intensive instruction following tasks | ROUGE L (%): 70.23 -> 70.23 | No clear effect |
| add reasoning layer on top of acting layer with memory (ablation) | MindExplore · 8 tasks, cruising and obstacle avoidance | average success rate (%): 60.00 -> 76.25 | Improved |
| add memory mechanism (ablation) | MindExplore · 8 tasks, cruising and obstacle avoidance | average success rate (%): 76.25 -> 92.50 | Improved |
| use point cloud instead of channel-concatenated 2D depth+RGB (ablation) | Base Expert · Moving to a Target Location | success rate: 32/50 -> 42/50 | Improved |
| add depth point cloud to RGB (ablation) | Arm Expert · Grasping a Specified Object | success rate: 23/50 -> 40/50 | Improved |
| add RGB to LiDAR+Depth point cloud (ablation) | Base Expert · Moving to a Target Location | success rate: 37/50 -> 42/50 | Improved |
| use expert-specific MMDP (BE/AE) instead of ACT (ablation) | acting layer · meta-action average | success rate (%): 36.86 -> 87.12 | Improved |
Closest papers by meaning
- ThinkAct: Vision-Language-Action Reasoning via Reinforced Visual Latent Planning (NeurIPS 2025 · similarity 0.92)
- VLA-OS: Structuring and Dissecting Planning Representations and Paradigms in Vision-Language-Action Models (NeurIPS 2025 · similarity 0.90)
- EfficientVLA: Training-Free Acceleration and Compression for Vision-Language-Action Models (NeurIPS 2025 · similarity 0.90)
- World–Value–Action Model: Implicit Planning for Vision–Language–Action Systems (NeurIPS 2026 · similarity 0.90)
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.90)
- CogVLA: Cognition-Aligned Vision-Language-Action Models via Instruction-Driven Routing & Sparsification (NeurIPS 2025 · similarity 0.90)
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.
