Offline Samples for Robotic Training
收藏arXiv2025-09-30 收录
下载链接：
https://sites.google.com/view/realrl
下载链接
链接失效反馈官方服务：
资源简介：
该数据集包含了40,000个离线样本，这些样本是通过400次运行收集而来的，旨在训练一个机器人代理，使其能够使用基于位置的可视伺服控制器和随机动作控制器来完成插孔任务。此外，收集这些样本的目的是为了在一个真实世界的机器人设置中评估反事实预测和各种奖励机制的有效性。该数据集的规模为40,000个样本，来自400次运行，任务重点是机器人强化学习在操作任务中的应用。
This dataset contains 40,000 offline samples collected from 400 trials, aimed at training a robot agent to complete peg-in-hole tasks using both position-based visual servo controllers and random action controllers. Additionally, these samples are collected to evaluate the effectiveness of counterfactual prediction and various reward mechanisms in a real-world robotic setup. The dataset consists of 40,000 samples from 400 trials, with the task focusing on the application of robotic reinforcement learning in manipulation tasks.
提供机构：
Research team conducting the robotic experiments
