isaaclab_ur7e_3task_fixed

Fixed LeRobot v3.0 datasets for the UR7e + RH5DG2 Isaac Lab tasks. TWO action conventions coexist, one folder set each; state / videos / prompts are identical between them, only action[:, 0:6] differs.

folder	arm action (dims 0..5)	hand action (dims 6..18)	eval execution
grasp_pan/, pour_cup/, cup_hang/	ABSOLUTE joint targets [rad] (= next recorded joint position)	default-relative offsets (default = 0)	ARM_ACTION=joint HAND_ABSOLUTE=0
grasp_pan_ee/, pour_cup_ee/, cup_hang_ee/	delta-EE of tool0 in the robot root frame, Diff-IK units: dpos/0.03 m, axis-angle drot/0.05 rad (rotation delta left-multiplied, Isaac Lab apply_delta_pose)	ABSOLUTE joint targets (= default 0 + offset, numerically identical)	ARM_ACTION=delta_ee_exec HAND_ABSOLUTE=1

Episodes: grasp_pan 199, pour_cup 150, cup_hang 200. Prompts: "Grasp the frying pan by its handle and place it on the nearest gas burner." / "Grasp the cup, lift it, and pour it out." / "Lift the hung cup off the holder".

Episode counts. grasp_pan: the phase file cook_food_dev2.hdf5 has 201 episodes; episode_0051 and episode_0097 ended without phase_3 (failed demos) and were excluded at render time, so the dataset has 199 complete demos (renumbered 0..198). pour_cup: 152 collected, episode_0105 / episode_0151 had no phase_3 and were removed, 150 remain. cup_hang: 200/200 complete. Train on all episodes -- do NOT pass --dataset.episodes. (The original pi05_grasp_pan run was launched with pour_cup's exclusion list [.. 105, 151 ..] by mistake, which dropped two valid grasp_pan episodes and listed two that do not exist; it trained on 197.)

What was wrong originally. DexSteer/isaaclab-ur7e-* have action[:, 0:6] == 0 in every frame (the phase collectors drove the arm with direct joint targets and recorded a zero arm action; isaac_tasks/data_collection/phase_collector.py, KNOWN DATA FLAW). Policies trained on them cannot move the arm. _pour_cup_clean_broken/ is the old pour_cup folder whose WRIST video came from the physics render and is corrupted (dark frame with a disc); pour_cup/ and pour_cup_ee/ carry the non-physics wrist render of the same recorded states.

How the EE labels were made. The recorded joint states were replayed in Isaac Lab and tool0 read back in the root frame (recover_arm_actions.py); the label is the delta between consecutive frames (re-applies through apply_delta_pose to 1e-6). Execution at eval (isaac_tasks/data_collection/ee_exec.py): integrate each delta on the COMMANDED tool0 pose, analytic UR IK seeded by the previous command (family-sticky, jump-limited), then the absolute JointPositionActionCfg arm term. Demo replay through that path: cup_hang 9/10, pour_cup 9/10 (same as joint-space). grasp_pan 4/10: its ready pose has wrist_2 = -pi (UR wrist singularity) at the start and at the placement, where any IK is ambiguous; the joint-space folder replays 10/10. Use the joint folders for grasp_pan unless the executor handles the singularity.

Evaluation env contract (isaac-tasks/scripts/pipeline/eval/run_pi05_ur7e.sh <task>): ID envs reproduce the collectors' physics -- grip friction 3.0 on robot + Cup (cup_hang also holder 0.08), hand gains 300/8/80 for pour_cup and grasp_pan, 450/8/130 for cup_hang, cup_hang firm arm gains 8000/400/400; cup_hang holder randomization = the training rectangle.

abs_ee variant (2026-09-09) — absolute EE pose, 6D rotation

grasp_pan_abs_ee/, pour_cup_abs_ee/, cup_hang_abs_ee/: action (22) = [ee_pos(3) | ee_rot6d(6) | hand(13)].

• ee_pos, ee_rot6d: ABSOLUTE tool0 pose (root frame) to reach at the NEXT frame; rotation is the 6D (first two columns of R) representation. hand = absolute joint targets (13).
• Why: the delta-EE (*_ee) target is 7-32x jumpier than joint targets (hard to learn) and its execution accumulates drift; the ABSOLUTE EE pose is 8-12x smoother AND its exact-label replay reaches 9-10/10 on all three tasks (grasp_pan included) because IK is re-seeded from the current measured joints each step (self-correcting, no drift).
• Eval: decode the predicted action to p=action[:3], R=decode_rot6d(action[3:9]) and drive the arm via ee_exec.ik_root(p, R, seed=current_joints) (the validated abs_ee executor). State/videos identical to *_ee.