Research map / Vision-language and multimodal / Robot action and navigation
SkillNet: Hierarchical Skill Modeling for Compositional Generalization in Vision-Language Action Models
Senwei Xie, Yuntian Zhang, Zhenzhou Tan, Ruiping Wang, Pengwei Wang, Shanghang Zhang, Xilin CHEN
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 skill-centric MoE (SkillNet) to pi0.5 | pi0.5 · LIBERO-Skill zero-shot | success rate: 31.8 -> 47.8 | Improved |
| replace vanilla MoE with SCMoE (ablation) | pi0.5 backbone · LIBERO-Skill zero-shot | success rate: 33.1 -> 47.8 | Improved |
| use vanilla MoE instead of nothing on pi0.5 | pi0.5 · LIBERO-Skill zero-shot | success rate: 31.8 -> 33.1 | No clear effect |
| add SkillNet to pi0.5 on real-world zero-shot | pi0.5 · Real-world compositional tasks | success rate: 42.5 -> 62.5 | Improved |
| add SkillNet few-shot transfer | pi0.5 · Real-world unseen behaviors 15-shot | success rate: +23.9% average | Improved |
| add SkillNet few-shot transfer | pi0.5 · RoboTwin-2.0 15-shot | success rate: +5.9% | Improved |
| add SkillNet in-domain | pi0.5 · LIBERO in-domain | success rate: 98.0 -> 99.1 | Improved |
| add SkillNet in-domain real-world long-horizon | Vanilla MoE · Real-world long-horizon | success rate: 50.0 -> 55.7 | Improved |
| add SkillNet on RoboCasa | pi0.5 and vanilla MoE · RoboCasa | success rate: +1.7% | Improved |
| add SCMoE instead of vanilla MoE (ablation) | SkillNet · Zero-shot generalization ablation | success rate | Improved |
| add hierarchical skill modeling (HSM) instead of flat skill IDs (ablation) | SkillNet · Few-shot transfer ablation | success rate | Improved |
| add router balance loss (ablation) | SkillNet · Few-shot transfer ablation | success rate: 6.2% degradation when removed | Improved |
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.
- AtomicVLA — AtomicVLA: Unlocking the Potential of Atomic Skill Learning in Robots (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
- Not All Features Are Created Equal: A Mechanistic Study of Vision-Language-Action Models (NeurIPS 2026 · similarity 0.89)
- VLA-OS: Structuring and Dissecting Planning Representations and Paradigms in Vision-Language-Action Models (NeurIPS 2025 · similarity 0.88)
- SemTra: A Semantic Skill Translator for Cross-Domain Zero-Shot Policy Adaptation (AAAI 2024 · similarity 0.87)
- Unveiling the Compositional Ability Gap in Vision-Language Reasoning Model (NeurIPS 2025 · similarity 0.87)
- Exploring the Limits of Vision-Language-Action Manipulation in Cross-task Generalization (NeurIPS 2025 · similarity 0.87)
- PriorVLA: Prior-Preserving Adaptation for Vision-Language-Action Models (NeurIPS 2026 · 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.
