Tag
Jim Fan's team open-sources V1.0 of ASPIRE, an agentic robotics framework with real robot skill discovery and self-improvement, noting frontier models like Astra, Opus 5.5, and Sonnet 5.5 have saturated their robot benchmarks.
提出GGSD框架，通过自博弈游戏引导技能发现，让人类可用少量离散技能直接控制具身智能体（Ant、Franka臂、Unitree G1），并在无需额外训练的情况下组合技能解决未见过的任务。
Video2Skill introduces a benchmark and the Streaming Embodied Skill Discovery (SESD) task, evaluating whether VLMs can organize streams of manipulation videos into reusable skills, revealing that model scale does not help and that recognizing when existing skills are insufficient is the central bottleneck.
DSD is a diffusion-based method for discovering diverse and reusable motor skills in simulated humanoid control, improving upon prior skill discovery techniques with broader behavioral coverage.
This paper presents a case study on skill discovery and routing in a multimodal agent harness, showing that partial in-prompt exposure of skills can create lexical competition that hinders correct selection, linking small-scale in-context retrieval to large-scale approaches.
ASPIRE is a continual learning system that autonomously develops and refines robot control programs through iterative exploration, achieving significant improvements in manipulation and household tasks while enabling sim-to-real transfer.
DataCOPE is an unsupervised verifier-guided skill discovery framework for data-analytic agents that derives verifier signals from exploration trajectories without labeled supervision. It improves performance by 9.71% and 32.30% on report-style and reasoning-style data analysis tasks respectively.
This paper introduces CARL, a method for offline hierarchical reinforcement learning that exploits local dynamics regularity to learn reusable skills. The approach clusters state-goal pairs requiring similar action sequences, enabling more effective skill reuse and improved performance on complex humanoid tasks.
SkillsVote is an AI Agent Skill management tool that filters 790K+ Skills from GitHub and extracts functional descriptions, environment requirements, and permission information. It supports precise recommendation, execution attribution, and iterative optimization, and can also provide workflow combination suggestions.
This paper proposes the Experience Compression Spectrum, a unifying framework that integrates agent memory, skill discovery, and rule-based systems along a single axis of increasing compression (5-20× for episodic memory, 50-500× for procedural skills, 1000×+ for declarative rules). The work identifies a critical gap—the 'missing diagonal'—showing that existing systems operate at fixed compression levels without adaptive cross-level support, and articulates design principles for scalable, full-spectrum agent learning systems.
SkillFlow introduces a benchmark of 166 tasks across 20 families for evaluating autonomous agents' ability to discover, repair, and maintain skills over time through a lifelong learning protocol. Experiments reveal a substantial capability gap among leading models, with Claude Opus 4.6 improving significantly while others show limited or negative gains from skill evolution.