Shiki42/ctr-pick-dual-bottles-uniform50-20260917

IdleMask review — passed (2026-09-19)

Reviewed by the dataset owner: `observation.arm_active_mask` is correct under the approved prefix-only rule. This section supersedes previous active/idle-mask descriptions below.

For each of the 50 episodes and each physical arm, only the initial contiguous wait before its first active sample may have mask 0. From the first active sample through the final frame, the mask is always 1, including intermediate waits, terminal holds, and repeated final actions. The initial activation boundary is preserved. Only arm_active_mask is authoritative; other idle or motion fields are not substitutes.

All 100 arm timelines passed binary and prefix-only validation. Actions, observations, timestamps, episode identities and videos are unchanged. Per-episode/global mask statistics and checksums were updated. See machine-readable review. The owner explicitly approved this repair and review status. This does not retroactively approve checkpoints trained on older revisions or their evaluation results.

pickdualbottles Uniform50

Open in LeRobot Dataset Visualizer

50 episodes from the same 50 source scene seeds, 11818 frames, 25 FPS, LeRobot v3, Aloha AgileX. Parent: Shiki42/ctr-pick-dual-bottles-uniform-20260916 at 9c6e24b90a0511b2f6d2021670870e6f9bdecd06 (repaired gripper target labels).

For source slots 0 through 49, select uniform0 at even slots and uniform1 at odd slots. The normalized relative-start positions are 0%, 2%, ..., 48%, 51%, 53%, ..., 99%. This is one sample per 2% bin, with 25 in each half. It is stratified approximate uniform coverage, not an exactly equally spaced 50-point grid. Uniformity refers to normalized nominal start offset, not actual overlap time. The mapping is in meta/uniform50-selection.json.

No scene replacement, new planning, simulation or trajectory generation. All selected numeric values are unchanged except global index and episode_index. Frozen control programs and actual source-index traces are retained. Scan synchronization and fixed tail, shoe mutual-exclusion waits, and idle masks remain as recorded. Action is the next sampled requested control target; observation.state retains the parent measured-state values. Full source-control gripper-label verification found zero stale initial-hold labels.

Three synchronized cameras, 320x240; both wrists retain centered_fovy90 and the calibrated centered mount. Selected videos are repacked into per-episode H264 High level3.0, yuv420p, CRF18, GOP25, 25 FPS with MP4 faststart. Frame counts, ordering and timestamps match the prior version; CRF18 uses normal lossy compression. No extra unselected episodes are included. Video encoding and decoded pixels change through standard CRF18 compression; all numeric trajectory rows remain byte-identical. All numeric and image statistics are recomputed on the selected frames. Image statistics use all RGB pixels with exact histogram quantiles, and numeric statistics use exact linear quantiles and population standard deviation.

Original joint names and ordering are retained in meta/info.json. Per-episode lineage, original camera/control receipts, and source-index traces are in meta/retime_manifest.json, meta/source_programs/ and meta/retime_source_indices/. Identity evidence is in .ctr-revision.json, meta/video-identity.json, meta/gripper-label-verification.json, and SHA256SUMS.

E232: CPU dataset subset construction and public publication. Status: reported, audit pending. Publication does not imply archival approval, a new physical success evaluation, or a learned-policy result. Parent collection and measured-state provenance limitations remain applicable. Training was not started by this release.

Video compatibility correction

This revision replaces the earlier CRF0 lossless encoding with the established CRF18 format for browser compatibility. Action, measured state, gripper labels, seed selection, masks and timing are unchanged. Image statistics were recomputed from the new decoded frames. Prior revision: 22eabe6e46fcf07f9db9070c3cabed9d0f2c662c.

Arm active mask

observation.arm_active_mask is float32[2], ordered left/right. At observation t it describes the outgoing action interval: 1 if the arm advances its frozen source program in any physical step of that interval. Gripper commands remain active even when joint positions do not change. Start delays, workspace waits, scan synchronization, cooperative holds and completion holds do not advance the program and are 0. This differs from both a motion threshold mask and the inverse of retime idle masks.

Intervals use the recorded samplephysicsstep boundaries (normally 10 physical steps at 250Hz / 25FPS), including a possibly shorter final interval. There is no extra terminal self-action row; the final actual action keeps its computed mask.

The algorithm follows the existing preview_mask.py logical-program-progress rule, generalized from fixed 15-step windows to recorded sample boundaries. Reference hash: db00172315f31f72fd0e7377a245e44266bdca9b57bc5e953b5a7bf2f2dc50de. Per-episode counts and trace identities are in meta/arm-mask-contract.json. All previous columns, videos and their CRF18 encoding remain unchanged. Source revision: c71b49c349f753ffed3e5890fb768ba96e700cd7.

Canonical dataset identity

This Uniform50 repository is the canonical identity for the single 50-episode / 11,818-frame bottles CTR dataset. Shiki42/ctr-pick-dual-bottles-ctr50-20260916 is a historical alias/source, not another sample set. Original and prefix-only corrected numerical columns were compared and are equal at their recorded source revisions. Do not concatenate the two names as 100 episodes. Historical Run/checkpoint bindings remain unchanged.