Takizz/piperx-water-delivery-0921-75ep-sequential
piperx-water-delivery-0921-75ep-sequential 150 episodes from 75 original sources, 171764 frames, 30 FPS. Three camera views and 14-dimensional action/state. Preparation ends at bottle opening +0.6s; synchronization ends at the later tray opening +0.2s. Both independent phases share one first arm and delay ratio. The shared tray motion remains synchronized. observation.arm_active_mask is float32[2], ordered left/right: 0 excludes an optional delay from supervision; 1 retains… See the full description on the dataset page: https://huggingface.co/datasets/Takizz/piperx-water-delivery-0921-75ep-sequential.
piperx-water-delivery-0921-75ep-sequential
150 episodes from 75 original sources, 171764 frames, 30 FPS. Three camera views and 14-dimensional action/state.
Preparation ends at bottle opening +0.6s; synchronization ends at the later tray opening +0.2s. Both independent phases share one first arm and delay ratio. The shared tray motion remains synchronized.
observation.armactivemask is float32[2], ordered left/right: 0 excludes an optional delay from supervision; 1 retains supervision, including required waits and completion. Raw samples are unchanged and have masks [1,1].
CTR uses four interleaved signed-offset samples per source. Sequential stores 75 left-first then 75 right-first episodes, both with p=1. Mixed stores left-first sources 1–50, right-first 26–75, then raw sources 1–25 and 51–75 (one-based IDs); each source occurs exactly twice. See provenance.json for the exact zero-based mapping.
Automated controls, masks, source coverage, video decoding and timestamps are validated. This is not exhaustive human visual acceptance; source segmentation/origin-audit warnings remain in the receipts. Split train/evaluation by original source identity, not by synthetic episode ID.