compsciencelab/BricksRL-Datasets

--- license: mit --- # BricksRL Dataset Card ## Dataset Summary The BricksRL dataset contains curated data for three robotic configurations: 2Wheeler, Walker, and RoboArm. The dataset includes expert and random data for four key tasks: Walker-v0, RoboArm-v0, RunAway-v0, and Spinning-v0. The expert data was collected using a trained Soft Actor-Critic (SAC) agent, while the random data was generated by executing a random policy. This dataset is presented in the paper [BricksRL: A Platform for Democratizing Robotics and Reinforcement Learning Research and Education with LEGO](https://arxiv.org/abs/2406.17490) (NeurIPS 2024). For more information feel free to check out the project [website](https://bricksrl.github.io/ProjectPage/). ## Supported Tasks The dataset supports the following tasks across various robot configurations: - Walker-v0 - RoboArm-v0 - RunAway-v0 - Spinning-v0 ## Dataset Structure The dataset contains two types of data: - Expert Data: Collected by a trained SAC agent solving the tasks on the real robot. The agent was evaluated over 100 episodes for each task, recording all transitions. - Random Data: Generated by executing a random policy on the real robot for 100 episodes per task. The datasets are TensorDicts, which can be directly loaded into the replay buffer. When initiating (pre-)training, provide the path to the desired TensorDict when prompted to load the replay buffer. Table 1 shows the dataset statistics regarding mean reward (expert data), number of transitions collected, and collection episodes. <div style="text-align: center;"> <img src="imgs/offline_dataset_stats.png" alt="stats" width="600"/> </div> ## Results and Evaluation The dataset was used to train both online and offline RL algorithms (Table 2). Performance comparisons between these methods demonstrated the effectiveness of offline RL algorithms, particularly when using expert data. Online RL algorithms struggled to generalize or often overfit when provided with expert demonstrations. For more detailed information about the hyperparameters, please refer to the appendix of the [paper](https://arxiv.org/abs/2406.17490). <div style="text-align: center;"> <img src="imgs/offline_results.png" alt="stats" width="600"/> </div> ## Citation If you use the BricksRL dataset in your research, please cite the following paper: ``` @article{dittert2024bricksrl, title={BricksRL: A Platform for Democratizing Robotics and Reinforcement Learning Research and Education with LEGO}, author={Sebastian Dittert and Vincent Moens and Gianni De Fabritiis}, journal={arXiv preprint arXiv:2406.17490}, year={2024} } ```--- license: mit --- # BricksRL 数据集卡片 ## 数据集概述 BricksRL 数据集包含针对三种机器人配置的精选数据：两轮机器人（2Wheeler）、步行机器人（Walker）和机械臂（RoboArm）。数据集涵盖四项核心任务的专家数据与随机数据：Walker-v0、RoboArm-v0、RunAway-v0 及 Spinning-v0。其中专家数据由训练完成的软演员评论家（Soft Actor-Critic, SAC）智能体采集，随机数据则通过执行随机策略生成。本数据集刊载于论文《BricksRL: A Platform for Democratizing Robotics and Reinforcement Learning Research and Education with LEGO》（NeurIPS 2024，arXiv:2406.17490）。更多信息可访问项目[官网](https://bricksrl.github.io/ProjectPage/)。 ## 支持的任务 本数据集支持以下跨多种机器人配置的任务： - Walker-v0 - RoboArm-v0 - RunAway-v0 - Spinning-v0 ## 数据集结构 本数据集包含两类数据： - 专家数据：由在实体机器人上完成任务求解的训练好的SAC智能体采集。该智能体针对每项任务进行了100个回合的评估，记录了所有状态转移样本。 - 随机数据：通过在实体机器人上执行随机策略，为每项任务采集100个回合的数据。 数据集采用TensorDict格式，可直接加载至经验回放缓冲区。在启动（预）训练时，加载经验回放缓冲区时请提供对应TensorDict的文件路径。 表1展示了数据集的统计信息，包括专家数据的平均奖励、采集的状态转移样本数量以及采集回合数。 <div style="text-align: center;"> <img src="imgs/offline_dataset_stats.png" alt="统计信息" width="600"/> </div> ## 结果与评估 本数据集被用于训练在线与离线强化学习算法（见表2）。不同方法的性能对比表明，离线强化学习算法具备有效性，尤其在使用专家数据时表现更优。在线强化学习算法难以实现泛化，或在获取专家演示后常出现过拟合现象。如需了解超参数的详细信息，请参阅该[论文](https://arxiv.org/abs/2406.17490)的附录。 <div style="text-align: center;"> <img src="imgs/offline_results.png" alt="结果" width="600"/> </div> ## 引用 若您在研究中使用BricksRL数据集，请引用以下论文： @article{dittert2024bricksrl, title={BricksRL: A Platform for Democratizing Robotics and Reinforcement Learning Research and Education with LEGO}, author={Sebastian Dittert and Vincent Moens and Gianni De Fabritiis}, journal={arXiv preprint arXiv:2406.17490}, year={2024} }