mbeukman/Kinetix-Offline

该数据集包含在Kinetix环境中收集的离线专家轨迹。Kinetix是一个基于JAX的2D刚体物理环境，其中的任务通过程序生成。每个任务共享相同的目标：使绿色和蓝色物体相互接触，同时避免绿色物体接触红色物体。智能体通过电机施加扭矩和推进器施加力来执行动作。针对每个程序生成的任务（即关卡）独立训练了专门的PPO代理，并且仅包含来自可解关卡的成功轨迹（约50%的生成关卡是可解的）。数据集总计包含来自超过1100万个独特任务的约30亿个状态转移。This dataset contains offline expert trajectories collected in Kinetix, a JAX-based 2D rigid-body physics environment where tasks are procedurally generated. Every task shares the same goal: make the green and blue objects touch, without green touching red. The agent acts by applying torques via motors and forces via thrusters. Specialist PPO agents were trained independently per procedurally generated task (i.e., level), and only successful trajectories from solvable levels are included (~50% of all generated levels are solvable). We have ~3B transitions from over 11M unique tasks.