RoboColosseum/g1-fold-towel

G1 Fold Towel

100 annotated fold towel demonstrations from robot Charlie, recorded on 2026-09-29. LeRobot v3.0, 30 FPS, 191,870 frames (1.777 hours).

The feature names, dimensions, camera layout and nested language schema follow BitRobot/HIW-500-LeRobot. These are separately converted source recordings, not copied reference episodes.

Features

• observation.images.head: 480 x 1280 RGB stereo image, kept side-by-side.
• observation.images.left_wrist, observation.images.right_wrist: 480 x 640 RGB.
• observation.state: 29 measured joint positions, ordered as listed in meta/info.json.
• observation.state.wbc: pivot (7), measured end-effector poses (12), gripper controls (4).
• action: pivot (7), commanded end-effector poses (12), gripper controls (4).
• language_persistent: main-task and subtask timeline with times relative to the episode start.
• language_events: empty lists, matching the reference dataset's representation.

The same pivot and gripper-control values accompany measured and commanded end-effector poses, following the reference's 23-dimensional layout. Head images are not split, and no camera flips or IR streams are added. Videos use AV1.

Segmentation and synchronization

Every exact fold towel annotation identifies an episode, closed by its explicit end. This release contains exactly 100 episodes, as requested. The existing 96 episodes are retained and four episodes are restored by trimming only the first tick (1/30 second) from each. Fresh source data arrives before the second tick, and all remaining samples pass the 2-second hold limit. The total removed duration is approximately 0.133 seconds. No interior gaps were joined and no end was moved. Seven other originally excluded intervals are not restored because the requested target is 100; their audit results and reasons are retained in meta/excluded_segments.json.

Video, measured joints, WBC state/action, and language times share the trimmed start. Original starts and annotations are preserved alongside effective starts in meta/source_segments.json and meta/trimmed_segments.json. Main-task and active subtask context at the new start are retained at relative time zero. Detailed before/after source timestamps are in meta/leading_gap_audit.json.

Each feature uses its most recent MCAP sample at or before each 30 Hz grid time, with a maximum hold of 2 seconds. No interpolation is applied. Existing 96 validated episodes were reused; only the four restored episodes were newly encoded. Encoding waits for every frame, avoiding frame drops. Original UTC timestamps are preserved without an additional timezone offset.

Language timelines are repeated on every frame of their episode, as in the reference. The full source annotation timeline is retained, including future subtasks; applications should filter by timestamp if they need only the current instruction. Calibration files are preserved under calibration/.

Validation and compatibility

All numeric frames were checked for finite values, correct dimensions, contiguous episode/frame indices and 30 Hz timestamps. First, middle and last frames of every episode were read back through the native LeRobot v3 reader and all three camera streams were decoded. See meta/validation.json.

The container uses LeRobot 0.4.4. Its native reader needs a feature/tensor adapter for the reference dataset's additive language type; the converter includes enable_hiw_language_reader() for that purpose. Numeric data, indexing and video decoding use the original LeRobot implementation. The exported Arrow language columns and feature declarations match the reference schema.