BASALT Evaluation and Demonstrations Dataset (BEDD)
收藏资源简介:
BEDD是由微软研究院等机构创建的大型数据集,包含2600万条图像-动作对,来源于近14000个视频,展示了人类玩家在Minecraft中完成BASALT任务的过程。数据集旨在用于训练和评估解决模糊任务的算法,特别是在缺乏明确奖励信号的情况下。BEDD不仅支持算法开发,还通过包含超过3000次密集的人类评估,为新算法提供了一个初步的基准。数据集的应用领域包括强化学习、模仿学习和人类反馈学习,旨在提高AI在复杂环境中的适应性和性能。
BEDD is a large-scale dataset created by Microsoft Research and other institutions. It consists of 26 million image-action pairs sourced from nearly 14,000 videos, which showcase the process of human players completing BASALT tasks in Minecraft. This dataset is intended for training and evaluating algorithms that solve ambiguous tasks, particularly in scenarios where explicit reward signals are absent. Beyond supporting algorithm development, BEDD also provides a preliminary benchmark for novel algorithms by incorporating over 3,000 dense human evaluations. The dataset has applications in reinforcement learning, imitation learning, and human feedback learning, with the goal of enhancing AI adaptability and performance in complex environments.

- 1BEDD: The MineRL BASALT Evaluation and Demonstrations Dataset for Training and Benchmarking Agents that Solve Fuzzy Tasks微软研究院 · 2023年



