[ICRL 2026 submission] SpoilDICE Experiment Dataset (1/2)

This content is shared anonymously for reviews and will include author and publication details if accepted. The dataset is organized into the following directory structure for 8 different tasks, SafetyAntVelocity-v1, SafetyHalfCheetahVelocity-v1, SafetyHopperVelocity-v1, SafetyWalker2dVelocity, SafetyPointCircle1-v0, SafetyPointCircle2-v0, ShadowHandOverSafeJoint, ShadowHandOverSafeFinger : The 'replay' directory contains replay data designed to simulate a natural data collection environment.It consists of three subsets: safe, performance, and shuffled. A detailed description of each dataset is provided in the Appendix of the paper. The 'controlled' directory includes datasets collected using four distinct types of policies: safe-high, safe-medium, unsafe-high, and unsafe-medium. These datasets capture different safety and performance trade-offs. For further explanations of each dataset, please refer to the Appendix of the paper. The 'stats' directory contains statistical information files summarizing all datasets within both the 'replay' and 'controlled' directories. These files report key statistics such as the mean return, mean cost, mean feasible return, and the number of trajectories, offering essential insights into the properties of each dataset.