Canonical dataset location
The maintained dataset is now published in the RSSR repository:
150case_data_enhance_blender_v2_trajectory_refine
That directory combines the previously selected original-v2 cases with newly refined trajectories. This standalone repository contains an earlier partial publication; its existing files are retained as a historical snapshot. New data and updates go to the canonical directory above.
S1 trajectory refinement — strict first-grasp cup stacking
Published: 20 / 38 currently admitted strict trajectories,
27391 synchronized frames, 60 videos. Origins:
1 refined, 19 original-clean.
All currently admitted sources published: False. The immutable initial refined snapshot contains 19 trajectories; its evidence is retained permanently.
The previous 150 attempted / 139 accepted background campaign is source provenance only;
those are not the counts of this dataset. Original-clean trajectories enter this repository only
after new equivalent full strict certification with fresh actual-force exact replay; heuristic-only
classification never suffices. Their original verified video and state/action bytes are reused.
Every immutable admission snapshot is retained under evidence/admission_snapshots/.
Refined sources take precedence for the same source group + case; the canonical dataset never duplicates a source.
Each source group and case appears once. Each cup has one meaningful full-episode closure,
with no regrasp, unplanned reopening, slip/drop, post-release reacquisition, or illegal release.
The first actual loaded bilateral acquisition remains stable until legal release, including
before the first 25 mm lift. Both cups pass the complete unchanged stacking task and its
hard constraints. Two independent replay paths are exact: ordinary native rollout/hard trace,
and a fresh actual-solver-load replay with exact saved rollout/hard trace. Force evidence
comes from actual solver state, not reconstruction without solver warmstart. Historical
quality_receipt.audit_criteria describes the earlier diagnostic before-analysis; authoritative
published force evidence is actual_force/receipt.json and native_replay_loads.npz.
All 21 mandatory quality checks and their immutable input hashes are included per episode.
The repository root is LeRobot v2.1: data/, videos/, meta/. Three synchronized RGB
cameras (head, left wrist, right wrist) use 224×224 H.264, 30 fps, with no image postprocessing.
One deterministic lighting augmentation is shared by all cameras and frames of each native
trajectory. Original rendered MP4 bytes are retained. Every video was decoded frame by frame,
including PTS and terminal-frame checks before publication.
observation.state: 32 measured scalar robot joints in compiled model order, retaining
all four torso joints, head, both arms, and all available gripper joints.
action: 22 original absolute desired joint targets in native actuator order, in radians.
Fixed-tendon gripper actions are the raw driver finger joint targets from issued commands;
they are not jaw aperture in metres or normalized 0–1 commands. Actuator force/ctrl is not action.
Names and exact state/action projection are in meta/info.json and each verification receipt.
Causal 30-Hz resampling never reads future commands and includes the terminal physical state.
Float32 Parquet values are verified against the original float64 arrays after the same cast;
aligned float64 NPZ and raw native proof arrays are retained under evidence/episodes/.
meta/source_cases.jsonl maps dataset episode to unique physical group + original source case
and labels origin explicitly. Group train/validation by source_case_id, across physical
variants; meta/source_case_groups.json supplies those groups. The default train split contains
all currently published episodes. Local source paths in receipts are provenance identifiers;
portable evidence links are relative to this repository. No model requests, environment logs,
or credentials are published. checksums.json covers every published file except itself.
Use LeRobot 0.3.3 / datasets 3.6.0 with PyAV:
from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("starry0929/v2-trajectory-refine", root="/path/to/downloaded/dataset", video_backend="pyav")
This is simulated S1 cup stacking data; hardware SDK calibration and real-robot success are not certified.
- Downloads last month
- -