自主学习技能基元库与技能库

本资源库主要面向智能体自主分层学习与技能迁移增强研究，针对动态不确定环境下复杂操作任务的策略复用需求建设。基于清华大学智能技术与系统国家重点实验室仿真-真机联动平台产生，主要记录了原子策略权重参数、128维技能嵌入向量、关键帧示例、状态机描述文件、行为树配置及仿真验证日志等核心内容。资源来源于在Unity-IsaacGym联动仿真平台中通过SAC/PPO算法进行的自监督强化学习训练，采用层次强化学习与蒙特卡洛树搜索方法实现技能基元的自动编排。产生方法包含三个阶段：首先每5000步训练保存策略与状态-动作对，收敛后筛选稳定基元并标注嵌入向量；然后基于层次化分解原理自动组合基元生成高阶技能；最后通过仿真安全验证与真机灰度测试的双重质控流程，确保技能可靠性与实用性。主要内容包含两个层次：技能基元层收录≥5000条原子策略，涵盖抓取、装配、定位等基础操作；组合技能层包含≈1500条由基元自动编排的高阶技能，支持复杂序贯任务的分解与执行。所有技能均采用JSON格式统一描述前置条件、动作轨迹与预期效果，重演验证成功率≥97%，技能标签经过两轮专家复核确保准确性。资源体量达2G，建设周期自2021年6月至2025年5月，遵循CC BY-NC 4.0许可协议，设置7年保存期限。本库通过提供可复用的策略基础与标准化的技能表示，有效突破智能体技能学习的策略增强瓶颈，为元学习方法在操作技能自主分层学习与迁移学习中的理论研究提供实践支撑，显著提升智能体在同类、异构、在线复杂操作任务中的适应能力与泛化性能。

This repository is primarily developed for research on autonomous hierarchical learning and skill transfer enhancement of AI Agents, targeting the strategy reuse requirement of complex manipulation tasks in dynamic and uncertain environments. Generated based on the simulation-real robot joint platform of the State Key Laboratory of Intelligent Technology and Systems, Tsinghua University, this dataset mainly records core contents including atomic policy weight parameters, 128-dimensional skill embedding vectors, key frame examples, state machine description files, behavior tree configurations, and simulation verification logs.

The dataset is derived from self-supervised reinforcement learning training using SAC/PPO algorithms on the Unity-IsaacGym co-simulation platform, where hierarchical reinforcement learning and Monte Carlo Tree Search (MCTS) are adopted to realize automatic sequencing of skill primitives. Its generation pipeline consists of three stages: First, save policies and state-action pairs every 5000 training steps, filter stable primitives and annotate their embedding vectors after training convergence; second, automatically combine primitives to generate high-level skills based on the hierarchical decomposition principle; finally, adopt dual quality control processes including simulation safety verification and real robot gray-scale testing to guarantee the reliability and practicality of the skills.

The main content is divided into two tiers: The skill primitive layer contains no fewer than 5000 atomic policies covering basic operations such as grasping, assembly, and positioning; the composite skill layer includes approximately 1500 high-level skills automatically sequenced by primitives, supporting decomposition and execution of complex sequential tasks. All skills uniformly describe preconditions, action trajectories, and expected outcomes in JSON format, with a replay verification success rate of no less than 97%. Skill labels have undergone two rounds of expert reviews to ensure their accuracy.

The total size of this repository is 2 GB, with a development cycle from June 2021 to May 2025. It follows the CC BY-NC 4.0 license and has a 7-year retention period.

This repository provides reusable policy foundations and standardized skill representations, effectively breaking through the strategy enhancement bottleneck in agent skill learning. It offers practical support for theoretical research on meta-learning methods applied to autonomous hierarchical learning and transfer learning of manipulation skills, and significantly improves the adaptability and generalization performance of AI Agents in homogeneous, heterogeneous, and online complex manipulation tasks.

搜集汇总

背景与挑战

背景概述

该数据集是一个面向智能体自主分层学习与技能迁移增强研究的资源库，包含5000多条原子策略和1500多条由基元自动编排的高阶技能，采用JSON格式统一描述，重演验证成功率≥97%。数据集体量为2.38GB，遵循CC BY-NC 4.0许可协议，旨在提升智能体在复杂操作任务中的适应能力与泛化性能。