Isaac Gym Environments

该数据集模拟了单臂或双臂机器人使用多指机械手进行灵巧操作任务的环境。通过基于种群的训练（PBT）算法评估性能和策略，显著提高了强化学习中的探索能力。在训练过程中，每个GPU上并行运行8192个环境，总计收集了0.32万亿次环境步骤。任务包括灵巧物体操作任务，如重新抓取、抓取并投掷以及物体重新定向。

This dataset simulates environments for dexterous manipulation tasks performed by single-arm or dual-arm robots equipped with multi-finger hands. Performance and policies are evaluated using the Population-Based Training (PBT) algorithm, which significantly enhances exploration capabilities in reinforcement learning. During training, 8192 environments run in parallel on each GPU, with a total of 0.32 trillion environment steps collected. The tasks include dexterous object manipulation tasks such as re-grasping, grasping and throwing, and object reorientation.