Codingpae/openarm_xemb_droid_pen_solverpath_droidgrasp_v2

OpenArm sim data: SolverPath-DroidGrasp, pen pick & place (2026-11-09)

Simulated OpenArm demonstrations. The grasp tilt and gripper close/open curves come from real DROID (Franka) demonstrations (lerobot/droid_1.0.1). The motion path is generated by a solver (waypoint IK + Lula RRT). Replayed with physics in an Isaac Sim 5.1 / Isaac Lab 2.3 digital twin and re-rendered with 3 cameras. Task: "pick up the blue pen and place it in the box".

• Format: LeRobot v3.0, 30 fps, observation.images.{headcam, camera_01, camera_02} (640×480), observation.state / action 16-D.
• 138 episodes, 102,434 frames = 108 first attempts (success and failure) + 30 successful retries.
• 39 of 147 source demos had no target grasp found on the first attempt (cause not yet determined).

Condition subsets (conditions_subsets.json)

Episode indices into this dataset.

subset	rule	episodes	frames
C1_none	all first attempts	108	79,180
C2_kin	first attempts passing the kinematic filter	108	79,180
C3_phys	first attempts that succeed in physics replay	92	66,979
C3r_retry	C3_phys + successful retries	119	87,895
C3r_clean	physics success and no arm contact during transit or after closure	114	84,558

Do not train on the full dataset with its meta/stats.json. It includes failed episodes. Build a subset dataset and recompute statistics first.

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 only; no real camera footage. Research use; please cite DROID. Contact: RV LAB.