File size: 2,410 Bytes
b7a9999 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 | ---
tags:
- robotics
- lerobot
- robotwin
- putcab
- preview
task_categories:
- robotics
---
# PutCab Concurrent: in-place waiting, Preview10
Ten qualified scene-matched demonstrations for user review. Three 320×240 H.264
camera views at 50/3 FPS; simulation runs at 250 Hz. Frames: 1641.
Left arm opens the drawer; right arm grasps and lifts the object at its original
location. Transfer and placement follow the required drawer readiness condition.
Sequential uses five A-then-B and five B-then-A demonstrations. Concurrent starts
both native expert programs at the same physics tick with independently advancing
control queues. Mixed selects five episodes from each parent on complementary
scene seeds. CTR uses the E098 V4 source trajectories, C2 transitions and selected
random deltas, with feedback waiting at the in-place safe-lift waypoint.
Ordinary expert episodes may use a front-clearance transit waypoint with
scene-specific approach height and orientation; each episode's transport strategy is recorded
in its lineage. Waiting remains at the original grasp-and-lift endpoint.
`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 task program
at any of the next 15 physics steps; 0 for pre-start, explicit holding, gate waiting,
completed programs and padding. Gripper commands and their programmed settling
dwell belong to the active program. The final repeated action has mask [0,0].
`observation.arm_motion_mask` separately reports whether any of the arm's seven
commanded coordinates changes by more than 1e-7 inside that same interval.
Measured physical joint motion is available in `meta/actual_joint_state`.
`meta/arm_progress` retains physics-step progress counters for independent mask
reconstruction. `meta/lineage.jsonl` records scene identities, command hashes,
mask receipts, scheduling events and parent revisions. These annotations are
part of the data artifact; model-side use is specified in separate training runs.
Each episode passed task success and full-scene penetration ≤10 mm in generation
and fresh exact-control video replay. All 30 videos were decoded and timestamps
checked; arm masks were reconstructed from the executed progress counters.
Experiment: E107-R001. Lifecycle: reported → audit pending.
Full 50-episode construction awaits user preview approval.
|