brandonyang/diffusion_policy_robocasa_activations_latest_chkpt

该数据集是从DiffusionTransformerHybridImagePolicy模型（使用diffusion_policy库）在RoboCasa基准任务上收集的每步、每集激活轨迹，采集于最新训练检查点。数据集包含7个RoboCasa任务（每个任务约30个回合，混合成功和失败执行），具体任务为：CloseFridge、CoffeeSetupMug、OpenDrawer、OpenStandMixerHead、PickPlaceCounterToCabinet、PickPlaceCounterToStove、TurnOnElectricKettle，并附带success_rates.json总结每个任务执行结果。数据总大小约327GB，文件结构按任务和步骤组织，关键文件包括suffix_residual.npz（用于概念器/SAE引导）、denoising.npz、adarms_cond.npz和奖励数据。该数据集旨在支持diffusion_policy实验中的概念器和稀疏自动编码器引导分析，适用于机器人学领域的扩散策略研究。This dataset contains per-step, per-episode activation traces collected from a DiffusionTransformerHybridImagePolicy model (from the diffusion_policy library) rolled out on RoboCasa benchmark tasks, captured at the latest training checkpoint. It includes 7 RoboCasa tasks (approximately 30 episodes each, with a mix of successful and failed rollouts): CloseFridge, CoffeeSetupMug, OpenDrawer, OpenStandMixerHead, PickPlaceCounterToCabinet, PickPlaceCounterToStove, and TurnOnElectricKettle, along with success_rates.json summarizing per-task outcomes. The total size is approximately 327 GB, organized in a hierarchical file structure with key files such as suffix_residual.npz (for conceptor/SAE steering), denoising.npz, adarms_cond.npz, and reward data. The dataset is designed for use with conceptor and sparse autoencoder steering pipelines in diffusion_policy experiments, supporting research in robotics and diffusion policy mechanisms.