HASARD
收藏资源简介:
HASARD是一个专为基于视觉的嵌入式安全强化学习设计的基准测试,由埃因霍温科技大学和利物浦大学共同开发。该数据集包含六个不同的随机环境,具有复杂的3D场景,要求安全策略和更高阶的推理。这些环境分为三个难度级别,并提供两种动作空间。HASARD的特点是模拟现实世界的导航和交互,同时保持计算效率。它旨在为视觉基础的安全RL研究提供一个低成本、高效的探索平台。
HASARD is a benchmark specifically designed for vision-based embedded safe reinforcement learning, co-developed by Eindhoven University of Technology and University of Liverpool. This dataset includes six distinct randomized environments with complex 3D scenes, which require safe policies and higher-order reasoning. These environments are divided into three difficulty levels and provide two action spaces. HASARD is characterized by simulating real-world navigation and interaction while maintaining computational efficiency. It aims to offer a low-cost, efficient exploration platform for vision-based safe RL research.




