Codingpae/openarm_xemb_droid_pen_trial5

OpenArm cross-embodiment sim data — DROID → OpenArm, pen pick & place (trial5, 2026-11-06)

Simulated OpenArm (bimanual, 7-DoF × 2, parallel gripper) demonstrations converted from real DROID (Franka) demonstrations (lerobot/droid_1.0.1), replayed with physics in an Isaac Sim 5.1 / Isaac Lab 2.3 digital twin of our real setup, and re-rendered with 3 cameras. Task: "pick up the blue pen and place it in the box" (identical string to our paired real dataset).

• Format: LeRobot v3.0, 30 fps, observation.images.{headcam, camera_01, camera_02} (640×480), observation.state / action 16-D (per arm: 7 joints [rad] + gripper motor angle [rad]) — same keys/units as the paired real teleop data.
• 161 episodes = 145 first attempts (success and failure, so baselines can be built from the same trajectories) + 16 successful retries.
• Wrist camera mount from CAD (hole-pattern registration) + Orbbec Gemini 305 datasheet (color origin = left RGB module) + factory intrinsics (HFOV 81.5°).
• Visual domain randomization (lighting, background, small camera pose jitter).

Condition subsets (conditions_subsets.json)

Episode indices into this dataset. All subsets come from the same converted trajectories — only the filter differs.

subset	rule	episodes
C1_none	all first attempts (no filter)	145
C2_kin	first attempts passing a kinematic filter (pose error / velocity limits, OXE-AugE / UMI style)	117
C3_phys	first attempts that succeed in physics replay (lifted, in box, released)	107
C3r_retry	C3_phys + episodes rescued by re-sampling the target grasp (up to 3 retries)	122

Physics verification and recording use the same start (rest pose → planned transit → replay); verdicts agree on 143/145 first attempts.

Provenance / license

Source trajectories: DROID (Khazatsky et al., 2024) via lerobot/droid_1.0.1 (listed as apache-2.0 on the Hub). Simulated renders contain no people. Research use; please cite DROID. Contact: RV LAB.