DenseReward dataset
收藏资源简介:
DenseReward数据集是由多机构联合构建的面向机器人操作任务的密集奖励标注数据集,包含27,000个轨迹片段。该数据集覆盖了从DROID、Isaac Sim、RoboSuite和LIBERO等多个平台采集的多样化场景和60多种操作物体,通过自动化数据生成管道合成了包括碰撞、抓取失败、物体掉落等六种物理真实失败模式。数据集创建过程基于五阶段操作分解框架,采用目标扰动方法在仿真环境中自动生成带有密集帧级奖励标签的轨迹,无需人工标注。该数据集主要应用于机器人强化学习领域,旨在解决稀疏奖励信号导致的信用分配问题,为视觉语言指导的机器人操作提供细粒度任务进度反馈。
The DenseReward dataset is a densely reward-annotated dataset for robotic manipulation tasks, jointly constructed by multiple institutions and containing 27,000 trajectory segments. This dataset covers diverse scenarios and over 60 manipulation objects collected from platforms including DROID, Isaac Sim, RoboSuite, and LIBERO. It synthesizes six physically realistic failure modes such as collision, grasp failure, and object dropping through an automated data generation pipeline. The dataset creation process is based on a five-stage manipulation decomposition framework, which uses the target perturbation method to automatically generate trajectories with dense frame-level reward annotations in simulation environments without manual annotation. This dataset is primarily applied in the field of robotic reinforcement learning, aiming to solve the credit assignment problem caused by sparse reward signals and provide fine-grained task progress feedback for vision-language guided robotic manipulation.

- 1DenseReward: Dense Reward Learning via Failure Synthesis for Robotic Manipulation北卡罗来纳大学教堂山分校; 卡内基梅隆大学; 上海交通大学; Amazon AWS AI · 2026年



