[Alpenj]

Scope

• derive the measured shoe-box pose from simulated H201 metric depth without MuJoCo body pose or segmentation IDs
• target the robot-facing front tuck flap right-hand grasp region
• solve guarded right-arm IK segments and add bounded right-wrist RGB correction
• keep the existing MuJoCo teleop, camera payload, episode quality gate, and DAPIER-native ACT dataset
• add a separate RoboTwin 2.0 Dual SO-101 embodiment and one deeply verified bimanual handover task

RoboTwin SIM result (2026-09-05)

• generated Dual SO-101 URDF from the official arm URDF and loaded 12 active joints in SAPIEN
• handled RTX 5050 / CuRobo 5-DoF IK without the failing fused LBFGS path
• split concave gripper collision meshes and added body, mast, opposite-arm collision geometry
• completed left grasp and lift, right approach, physical transfer, gradual left release, and left retreat
• final gate: both right jaws in contact, opposing normals -0.96178, inter-arm impulse 0, object z 0.899101 m
• no equality constraint or fake object attachment
• generated H.264 video with 1,256 frames and HDF5 with 1,255 state/action steps
• converted the real episode to 12D state/action, left/right 320x240 RGB, and H201 640x460 uint16-mm depth
• replaced broken OIDN ray-traced output with a DAPIER-only SAPIEN raster path
• record measured SAPIEN qpos instead of drive targets; maximum 15 Hz arm step is 0.033236 rad

Verification

• DAPIER RoboTwin 12-axis and mount contract: PASS
• RoboTwin HDF5 to canonical episode fixture: PASS, including unsafe action rejection
• official runtime installer and SAPIEN load: PASS, 12 joints / 25 links
• successful physical-contact SIM seed: 3; failed seeds 0-2 were not counted
• actual HDF5 canonical conversion gate: PASS
• Python compilation, patch reverse-check, and git diff check: PASS

Not yet complete

• multiple-seed robustness and measured-range domain randomization
• ACT training and held-out RoboTwin evaluation with these episodes
• physical wrist/H201 intrinsics and extrinsics
• STS3215 dynamics identification, follower calibration mapping, shadow inference, and low-speed rollout
• physical lid opening and shoe extraction

This PR remains Draft. SIM and mock evidence does not authorize physical execution.