Skill Discovery | 无监督技能发现的经典工作总结

目录

• [🐱 Unsupervised](#🐱 Unsupervised)
  • [Diversity is All You Need: Learning Skills without a Reward Function (diayn)](#Diversity is All You Need: Learning Skills without a Reward Function (diayn))
  • [Explore, Discover and Learn: Unsupervised Discovery of State-Covering Skills (EDL)](#Explore, Discover and Learn: Unsupervised Discovery of State-Covering Skills (EDL))
  • [CIC: Contrastive Intrinsic Control for Unsupervised Skill Discovery](#CIC: Contrastive Intrinsic Control for Unsupervised Skill Discovery)
  • [Lipschitz-constrained Unsupervised Skill Discovery (LSD)](#Lipschitz-constrained Unsupervised Skill Discovery (LSD))
  • [Controllability-Aware Unsupervised Skill Discovery (CSD)](#Controllability-Aware Unsupervised Skill Discovery (CSD))
  • [METRA: Scalable Unsupervised RL with Metric-Aware Abstraction](#METRA: Scalable Unsupervised RL with Metric-Aware Abstraction)
  • [Can a MISL Fly? Analysis and Ingredients for Mutual Information Skill Learning (csf)](#Can a MISL Fly? Analysis and Ingredients for Mutual Information Skill Learning (csf))
  • [Foundation policies with hilbert representations (HILP, offline metra)](#Foundation policies with hilbert representations (HILP, offline metra))
  • [Disentangled Unsupervised Skill Discovery for Efficient Hierarchical Reinforcement Learning (DUDSi)](#Disentangled Unsupervised Skill Discovery for Efficient Hierarchical Reinforcement Learning (DUDSi))
  • [SkiLD: Unsupervised Skill Discovery Guided by Factor Interactions](#SkiLD: Unsupervised Skill Discovery Guided by Factor Interactions)
  • [Efficient Skill Discovery via Regret-Aware Optimization](#Efficient Skill Discovery via Regret-Aware Optimization)
• [🦜 Guided](#🦜 Guided)
  • [Safety-Aware Unsupervised Skill Discovery](#Safety-Aware Unsupervised Skill Discovery)
  • [Do's and Don'ts: Learning Desirable Skills with Instruction Videos (dodont)](#Do's and Don'ts: Learning Desirable Skills with Instruction Videos (dodont))
  • [Language Guided Skill Discovery (LGSD)](#Language Guided Skill Discovery (LGSD))
  • [Reference Guided Skill Discovery (RGSD)](#Reference Guided Skill Discovery (RGSD))
  • [Controlled Diversity with Preference: Towards Learning a Diverse Set of Desired Skills (CDP)](#Controlled Diversity with Preference: Towards Learning a Diverse Set of Desired Skills (CDP))
  • [Human-Aligned Skill Discovery Balancing Behaviour Exploration and Alignment (HaSD)](#Human-Aligned Skill Discovery Balancing Behaviour Exploration and Alignment (HaSD))
  • [Guiding Skill Discovery with Foundation Models (fog)](#Guiding Skill Discovery with Foundation Models (fog))

---

🐱 Unsupervised

Diversity is All You Need: Learning Skills without a Reward Function (diayn)

• ICLR 2019。
• arxiv：https://arxiv.org/abs/1802.06070
• pdf：https://arxiv.org/pdf/1802.06070
• html：https://ar5iv.labs.arxiv.org/html/1802.06070
• website：https://sites.google.com/view/diayn
• 博客：论文速读纪录 | 2025.01

Explore, Discover and Learn: Unsupervised Discovery of State-Covering Skills (EDL)

• ICML 2020。
• arxiv：https://arxiv.org/abs/2002.03647
• GitHub：https://github.com/victorcampos7/edl
• 博客：论文速读记录 | 2026.02

CIC: Contrastive Intrinsic Control for Unsupervised Skill Discovery

• 疑似 ICML 2022。
• arxiv：https://arxiv.org/abs/2202.00161
• pdf：https://arxiv.org/pdf/2202.00161
• html：https://ar5iv.labs.arxiv.org/html/2202.00161
• website：https://sites.google.com/view/cicrl/
• GitHub：https://github.com/rll-research/cic
• 博客：论文速读纪录 | 2025.01

Lipschitz-constrained Unsupervised Skill Discovery (LSD)

• ICLR 2022。
• arxiv：https://arxiv.org/abs/2202.00914
• pdf：https://arxiv.org/pdf/2202.00914
• html：https://ar5iv.labs.arxiv.org/html/2202.00914
• 博客：论文速读记录 | 2025.12（1）

Controllability-Aware Unsupervised Skill Discovery (CSD)

• ICML 2023。
• arxiv：https://arxiv.org/abs/2302.05103
• pdf：https://arxiv.org/pdf/2302.05103
• html：https://ar5iv.labs.arxiv.org/html/2302.05103
• 博客：论文速读记录 | 2025.12（1）

METRA: Scalable Unsupervised RL with Metric-Aware Abstraction

• ICLR 2024 Oral。
• arxiv：https://arxiv.org/abs/2310.08887
• pdf：https://arxiv.org/pdf/2310.08887
• html：https://arxiv.org/html/2310.08887
• website：https://seohong.me/projects/metra/
• GitHub：https://github.com/seohongpark/METRA
• 博客：Skill Discovery | METRA：让策略探索 state 的紧凑 embedding space

Can a MISL Fly? Analysis and Ingredients for Mutual Information Skill Learning (csf)

• ICLR 2025 oral。
• arxiv：https://arxiv.org/abs/2412.08021
• open review：https://openreview.net/forum?id=xoIeVdFO7U
• GitHub：https://github.com/Princeton-RL/contrastive-successor-features
• 博客：论文速读记录 | 2025.06

Foundation policies with hilbert representations (HILP, offline metra)

• ICML 2024。
• arxiv：https://arxiv.org/abs/2402.15567
• website：https://seohong.me/projects/hilp/
• 博客：论文速读记录 | 2025.12（2）

Disentangled Unsupervised Skill Discovery for Efficient Hierarchical Reinforcement Learning (DUDSi)

• neurips 2024，7 5 5 4 poster。
• arxiv：https://arxiv.org/abs/2410.11251
• open review：https://openreview.net/forum?id=ePOBcWfNFC
• website：https://jiahenghu.github.io/DUSDi-site/
• 博客：论文速读记录 | 2025.09

SkiLD: Unsupervised Skill Discovery Guided by Factor Interactions

• neurips 2024，8 6 5 5 poster，5 是 borderline ac。
• arxiv：https://arxiv.org/abs/2410.18416
• open review：https://openreview.net/forum?id=i816TeqgVh
• website：https://wangzizhao.github.io/SkiLD/
• 博客：论文速读记录 | 2025.09

Efficient Skill Discovery via Regret-Aware Optimization

• ICML 2025，3 3 2 1 poster。
• arxiv：https://arxiv.org/abs/2506.21044
• open review：https://openreview.net/forum?id=4qMJ8Ignmp
• GitHub：https://github.com/ZhHe11/RSD
• 博客：论文速读记录 | 2025.10

🦜 Guided

Safety-Aware Unsupervised Skill Discovery

• ICRA 2023。
• paper：https://safe-skill.github.io/static/pdfs/safe-skill.pdf
• website：https://safe-skill.github.io/
• 博客：CSDN |【ICRA 2023】SASD 论文阅读笔记：一种安全感知的无监督技能发现方法

Do's and Don'ts: Learning Desirable Skills with Instruction Videos (dodont)

• NeurIPS 2024 poster。
• arxiv：https://arxiv.org/abs/2406.00324
• pdf：https://arxiv.org/pdf/2406.00324
• html：https://arxiv.org/html/2406.00324
• website：https://mynsng.github.io/dodont/
• open review：https://openreview.net/forum?id=7X5zu6GIuW
• 博客：Skill Discovery | DoDont：使用 do + don't 示例视频，引导 agent 学习人类期望的 skill

Language Guided Skill Discovery (LGSD)

• ICLR 2025，8 8 6 6 poster。
• arxiv：https://arxiv.org/abs/2406.06615
• pdf：https://arxiv.org/pdf/2406.06615
• html：https://arxiv.org/html/2406.06615v2
• open review：https://openreview.net/forum?id=i3e92uSZCp
• 博客：Skill Discovery | LGSD：用描述 state 的语言 embedding 的距离，作为 metra 的 d(x,y) 距离约束

Reference Guided Skill Discovery (RGSD)

• ICLR 2026。
• arxiv：https://arxiv.org/abs/2510.06203
• pdf：https://arxiv.org/pdf/2510.06203
• html：https://arxiv.org/html/2510.06203
• open review：https://openreview.net/forum?id=IaGf8Eh5Uo
• 博客：Skill Discovery | Skill Discovery | RGSD：基于高质量参考轨迹，预训练 skill space

Controlled Diversity with Preference: Towards Learning a Diverse Set of Desired Skills (CDP)

• AAMAS 2023。
• arxiv：https://arxiv.org/abs/2303.04592
• GitHub：https://github.com/HussonnoisMaxence/CDP
• 期刊版本：Human-informed skill discovery: Controlled diversity with preference in reinforcement learning，science direct。
• 博客：论文速读记录 | 2025.11

Human-Aligned Skill Discovery Balancing Behaviour Exploration and Alignment (HaSD)

• AAMAS 2025。
• arxiv：https://arxiv.org/abs/2501.17431
• GitHub：https://github.com/HussonnoisMaxence/HaSD-AAMAS
• 博客：论文速读记录 | 2025.11

Guiding Skill Discovery with Foundation Models (fog)

• 最新论文链接：https://liacs.leidenuniv.nl/~plaata1/papers/4848.pdf
• ICLR 2025 版 open review 论文链接：https://openreview.net/pdf?id=nZBUtzJhf8
• 最新 website：https://sites.google.com/view/submission-fog （可惜有一些可视化好像挂掉了）
• 博客：Skill Discovery | FoG：使用 LLM / CLIP 给出 dodont 权重，以引导 agent 安全探索