Episodes Preview SO-101 Visualizer
120 episodes · 20 fps · 3 cameras · 640×480 av1

SO-101 Lab01 Pick-and-Place Dataset (yaw-aware agent, v3)

Privileged-state scripted expert recorded with SO-101 SimStudio into LeRobot v3.0. Built with the post-2026-09-28 expert fixes (wrist roll aligned to the approach face, two-jaw contact gate, far-reach quality rejects). Does not overwrite v1, v2, or v2-clean.

Hub id alexhegit/so101-simstudio-lab01-pnp-agent-yaw-v3
Robot/action so101_mujoco, 6-D joint position
State 15-D position + velocity + EE XYZ
EE point fingertip site gripperframe
Cameras front, top, wrist (640×480, 20 Hz)
Spawn x [0.26, 0.34], y [0.165, 0.235], yaw ±45°
Size 120 episodes / 52476 frames

Composition

Kind Count Notes
First-try success 100 Dual quota: 6 yaw bins × 2×2 XY cells (25 each); one grasp from home
Pre-close correction 10 Gripper still open; small align, then one grasp; full home→place
Partial-miss local reclose 10 Half-open (+0.50) miss not labeled; open in place and regrasp (no high retreat)

Deliberate wrong actions are never saved as expert labels. Failed attempts are discarded; temporary video PNGs are cleared so episode videos match parquet length (no trailing tails).

Expert fixes vs v1 / v2

  • Wrist roll chosen so the jaw axis is perpendicular to the approach-face normal (square 90° ambiguity removed).
  • Approach face ranked by roll cost, then standoff reach.
  • Close rejected unless both jaws contact the block (≥1 N) and pre-close tangent error ≤10 mm.
  • Far-table unreachable poses are quality-rejected, not saved as empty grasps.

Collector: labs/lab01_pnp/auto_collect_yaw.py. spawns.jsonl records spawn, yaw bin, XY cell, recovery style, and success.

Downloads last month
137