Piper X General Pickup — Same-cluster preview
4 successful episodes, 694 frames, 25 FPS; three synchronized 640×480 RGB cameras.
这一版将目标和干扰物放在同一簇内，取消干扰物的前后分区以及目标周围 7.5 cm 的统一禁入区。 至少三个干扰物初始摆在目标 XY 包围盒附近，初始间距为 1.5–3 cm；其余物体沿同一簇扩展。 物理稳定后与录制开始时，均要求至少三个刚体干扰物到目标的 XY 包围盒间距不超过 6 cm。 这些距离为保守包围盒足迹代理，不能视为精确网格表面距离或遮挡指标。 本预览为近邻、不重叠的桌面布局；不包含有意堆叠。
The target is physically grasped near its live mesh centre, approached and descended to with open fingers, then closed upon only after measured alignment passes. The inactive arm stays at joint-zero Home. The target is lifted approximately 10 cm and held for one second, with no attachment or teleport during recording. Existing collision planning remains enabled. The four successful layouts are replayed from a first collection that needed 28 attempts. This replay uses their final seeds and is not an independent success-rate estimate. This small preview does not establish general clutter robustness.
| Episode | Target | Frames | Near neighbours (≤6 cm AABB gap) | Closest AABB gap (cm) | Lift (cm) |
|---|---|---|---|---|---|
| 0 | correction_tape/00000 | 176 | 3 | 1.64 | 9.98 |
| 1 | sachet/00000 | 174 | 4 | 2.07 | 9.98 |
| 2 | correction_tape/00000 | 170 | 5 | 1.41 | 9.98 |
| 3 | chocolate_bar/00000 | 174 | 3 | 0.35 | 9.98 |
Format: LeRobot v3.0. Cameras: observation.images.top, left_wrist, right_wrist.
observation.state and action: selected arm's six joint positions/targets (radians), plus gripper-open fraction.
observation.active_arm: left=0, right=1. observation.environment_id: workspace index.
annotation.phase: initial=0, approach=1, descend=2, close=3, lift=4, hold=5.
All raw camera videos decoded fully with aligned frame counts. Export used the official LeRobot v3 writer;
first/middle/last frames in every episode were read with the official reader.
Provenance, layout, retries and measured neighbourhoods: meta/a2r_manifest.json.
Collector code: Shiki42/a2r PR #1.
Render quality correction
The current revision replaces the prior preview that flickered because DLSS history was cleared per capture.
RTX accumulation and a one-shot DLSS reset are applied only at the wave boundary, including after eight renders that synchronize new poses/materials; temporal history is retained throughout recording. Quality mode and two delivered render passes
per captured physics state retain 25 FPS and 640×480 images. Initial/pre-close frames in all four episodes
were visually reviewed across two collection waves. Raw static-wall frame differences were also measured in initial and hold phases; see meta/temporal_quality.json. Initial layouts / Pre-close layouts.
Before / after video: same target seed, left prior flickering version, right current capture.
- Downloads last month
- 706