Episodes Preview dobot_cr3a Visualizer
204 episodes · 30 fps · 1 camera · 640×480 h264

WMX R2 CR3A — camera_30fps

Public cumulative release: 204 recorded runs = the original 59 + 145 audited additions. LeRobot v3.0; 100,894 frames; nominal 30 Hz. These runs are not 204 confirmed successful tasks. Both repositories represent the same source cohort.

이 저장소는 영상과 축 데이터를 고정 30 Hz 격자에 맞춰 선택한 카메라 파생본입니다. 축 로그의 전체 1 kHz 샘플이 들어 있는 저장소는 아래 연결된 Axis입니다.

대응 데이터셋: movensys/wmx-r2-command-grounded-cr3a-axis-v0.

Camera 데이터 구성과 시간축

이 데이터셋은 204개 실행의 손목 RGB 영상과 축 명령·피드백을 고정 30 Hz 격자에 맞춘 100,894개 샘플로 제공합니다. 영상 기반 학습과 영상·축 상태의 함께 보기에 사용합니다.

원본 축 데이터의 취득과 30 Hz 변환

원본 축 로그는 WMX3의 EtherCAT DC 동기화된 1 kHz 제어 주기에서 취득했습니다. Camera 데이터셋에서는 이 원본 로그의 모든 행을 사용하는 대신, 30 Hz 격자 시점마다 최근접 축 로그 행을 선택합니다. 원본 제어 주기의 DC 동기화와 영상·축 데이터의 후처리 시간 대응은 별개입니다. 원본 축 로그가 DC 동기화돼 있다는 사실이 카메라 노출도 DC에 동기화됐다는 뜻은 아닙니다.

영상과 축 데이터 선택 방법

  1. 각 실행의 첫 축 호스트 시각부터 마지막 축 호스트 시각 이내까지 고정 30 Hz 격자(약 33.333 ms 간격)를 생성합니다.
  2. 각 격자 시점에 가장 가까운 카메라 호스트 수신 시각의 RGB 프레임을 선택합니다.
  3. 같은 격자 시점에 가장 가까운 WMX 호스트 시각의 원본 축 로그 행을 선택하고 그 행의 명령·피드백·추종 오차를 사용합니다.
  4. 거리가 같은 두 후보 중에는 앞선 값을 선택합니다. 수치 평균·보간·저역통과 필터는 적용하지 않습니다. 학습용 수치 특성은 float32로 표현합니다.

실제 카메라 노출 시각을 기준으로 축 로그를 직접 매칭한 것이 아니라, 고정 30 Hz 격자에 영상과 축 데이터를 각각 최근접 선택한 방식입니다. 원본 RGB 프레임이 반복 선택되거나 일부 프레임이 선택되지 않을 수 있으며, 에피소드별 선택 집계는 annotations/episode_mapping.json에 있습니다. 전체 1 kHz 신호의 고주파 변화나 짧은 이벤트 분석에는 원본 해상도의 축 데이터를 사용해야 합니다.

Timestamp와 원본 시각의 구분

LeRobot timestamp = frame_index / 30은 에피소드 시작을 0초로 하는 상대 시간입니다. 약 33.333 ms 간격의 샘플 순서를 표현하며, DC 절대 시각 또는 실제 카메라 노출 시각을 직접 저장한 값이 아닙니다.

필드 의미
timestamp 에피소드 내 frame_index / 30, 초 단위 상대 시간
source.grid_host_time_ns 영상과 축 로그를 선택할 때 사용한 30 Hz 격자 시각
source.axis_host_time_ns 선택한 원본 축 행의 SYSTEM_TIME × 100, 정수 나노초 호스트 시각
source.camera_host_time_ns 선택한 RGB 프레임의 호스트 수신 시각
source.camera_header_ns 선택한 RGB 프레임의 ROS 헤더 시각
source.cycle 및 source.axis_row 선택한 축 행의 원본 엔진 사이클과 행 번호
source.camera_frame 선택한 원본 영상 프레임 번호
source.camera_minus_grid_ns 선택한 프레임의 수신 시각 − 격자 시각
source.axis_minus_grid_ns 선택한 축 행의 호스트 시각 − 격자 시각

원본 TIME 및 SYSTEM_TIME은 raw 패키지에 보존합니다. SYSTEM_TIME을 DC 시계의 원본 값으로 간주하지 않습니다. 영상 수신 시각과 축 호스트 시각의 차이는 위 두 offset 열의 차로 계산할 수 있지만, 센서 노출과 제어 신호의 물리적 동기화 오차를 의미하지는 않습니다. 카메라 노출·ROS 헤더·호스트 수신 시각과 WMX/DC 시계 사이의 정밀 정합은 이번 패키징 검증의 완료 항목이 아닙니다.

English summary

The source axis logs were acquired during WMX3 EtherCAT DC-synchronized 1 kHz control cycles. This camera dataset selects one nearest source axis row and one nearest RGB frame independently for each point on a uniform 30 Hz host-time grid. RGB selection uses host reception time, not a calibrated exposure timestamp. No interpolation, averaging or anti-alias filter is applied; numeric model features are float32. RGB source frames may repeat or be omitted.

LeRobot timestamp = frame_index / 30 is relative time in seconds, starting at zero per episode. It is neither an absolute DC clock value nor the actual exposure timestamp. Original cycle numbers, row/frame indices, host timestamps, ROS header timestamps and grid offsets remain separate provenance fields. DC-synchronized axis acquisition does not establish DC synchronization of the camera. Physical exposure-to-control clock alignment has not been validated by this packaging process.

Loading

Tested with Python 3.12, lerobot==0.4.4 and av==15.1.0; see requirements-lock.txt for the complete environment.

from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("movensys/wmx-r2-command-grounded-cr3a-camera-v0", revision="cumulative-204-language-v1-20261102", video_backend="pyav")
sample = ds[0]

Initial download requires network access. No robot commands are issued. Keep meta/, data/ and videos/ together. For a local download, pass root="/absolute/path/to/download".

Original data, labels and examples

Original CSV/AVI, timestamps and raw examples are stored once in the axis repository raw package. Raw package paths in annotations resolve relative to that directory. annotations/ contains integrated labels and episode_mapping.json; original source paths are provenance identifiers. Use examples/plot_dataset.py with --root pointing to this downloaded repository and --episode 162 --output example.png. plots/ contains executed examples. The bundled load_local.py and validate_bundle.py target the full local bundle layout; the direct loading example above targets this standalone repository.

Feature meanings

action: six command positions then six command velocities (configured rad, rad/s). observation.state: six feedback positions then six feedback velocities. observation.tracking_error: command-minus-feedback positions computed before float32 conversion. RGB loads as (3,480,640) float values in [0,1]. Engine setpoints are not a complete deployed policy interface. Source indices and timestamps remain int64; raw numeric data stays float64. FBTRQ is excluded from learning features because ENI mapping is unverified. No measured gripper force/position or continuous depth is claimed.

Timing and quality

Camera grid samples use nearest RGB reception timestamp and nearest axis host timestamp, ties earlier; there is no numeric interpolation. Repeated/omitted RGB selections and maximum offsets are stored per episode. H.264/yuv420p is lossy. The additions retain 22 reception-gap alerts over 50 ms; these do not prove missing exposures. The source axis logs were acquired in DC-synchronized 1 kHz control cycles; the cross-clock mapping between camera exposure/header, host receipt and WMX/DC time has not been validated. Nearest-time residuals are not physical synchronization accuracy. Separate runs and preflight streams are not continuous task videos.

Labels and scope

The baseline59 remains a fixed cohort. Additional T008 correction runs are now included separately and do not overwrite T007's partial orientation outcome. Incoming corrections and evidence addenda are captured in the integrated label overlay. Unknown physical outcomes remain unreviewed; controller success, user confirmation and image estimates are separate. Missing task links remain null; UNASSIGNED per-run lookup keys are placeholders, not new confirmed task chains. Consult annotations/integration_notes.json, annotations/camera_gap_review.json, and episode applied_addenda. Labels are not a newly adjudicated physical-success ground truth. Original source ledgers were not changed.

No held-out split is supplied. Group related work by task/session; the default LeRobot train split is storage only. Original source_path strings remain provenance identifiers; use package paths and episode_mapping for loading. Failure/multistream cases not selected for the 145 additions remain outside this converted cohort.

Versions and license

This release replaces the default 59-run snapshot with the cumulative 204-run snapshot; do not append all 204 to the old 59. The former version is preserved as baseline-59-before-20261102. Pin revision="cumulative-204-language-v1-20261102" for reproducibility. Aggregate statistics were recomputed over 204 runs. No explicit open-data license is assigned; public visibility does not itself grant an open license.

Language instruction correction on 2026-11-02

135 added episodes (134 distinct task strings) previously used task IDs as their descriptive prefix. They now have Korean descriptions of intended operations in both meta/tasks.parquet and meta/episodes. These descriptions are based on the recorded operations, not new physical-success judgments. The original 59 episode instructions, frame task_index values, numeric data, videos, task identities and 35 null task links are unchanged. annotations/language_instruction_overrides.json records the before/after wording and evidence basis. The prior cumulative-204-20261102 tag remains immutable. See language_instruction_validation.json for per-episode official-loader checks.

Downloads last month
243