kt_pick_pringles_spam_modular_merge_mujoco

MuJoCo real2sim re-log of learner1119/kt_pick_pringles_spam_modular_merge: every recorded episode replayed with its recorded actions in a calibrated MuJoCo scene of the same robot (FFW SH5), table, head camera and objects (YCB 001 chips can, 010 potted meat can). Same LeRobot v2.1 layout, keys, shapes, fps and episode/task tables as the source.

Key	Content
observation.state (57)	simulated joint positions, source joint order; frame t is the state before action t
action (57)	source action, unchanged
observation.tactile (90)	copied from the source: the simulation has no tactile sensor
observation.images.cam_head	simulated head camera (ZED Mini left, 672 x 376, 10 fps), rendered with MuJoCo
timestamp, frame_index, episode_index, index, task_index	as in the source

Episodes: 195, frames: 24052. Median joint rms, sim state vs. recorded state: 0.0335 rad.

Simulation

• MuJoCo 3.11, timestep 0.25 ms. The robot base is pinned at the same world pose in every episode (0.4, 0, 0.0035 m, parallel to the table); the per-session robot-table offset, fitted from the head camera table silhouette, moves the table (table_pos in meta/real2sim_episodes.jsonl), and the table height per session is bounded by the hand-table clearance. Head (tilt 37.2 deg, pan 0) and lift (-0.130 m) are held at their command, as on the real robot. Head camera at the URDF optical frame with a fitted tilt/pan correction.
• Command transport fitted on the right arm before grasp (Stage A): 61.4 ms delay, first-order lag tau 89.1 ms, right-arm servo gain x5.89 and damping x0.89 sqrt(5.89), applied to all 57 commands (gain scale on the right arm only).
• Object start poses from a frame-0 silhouette fit of the source image (poses_image). Collision: cylinder (chips can, 205 g) and box (meat can, 370 g). Right hand-object contacts use explicit pairs (condim 4, friction 1.0, torsion 0.10, solref 5 ms, solreffriction 0.5 ms).
• Rendering only: one spot light above and behind the head, weak headlight, YCB textures with value/saturation gains, table colour matched to the real table.

Per-episode sim labels: meta/real2sim_episodes.jsonl

sim_success = 1 when the target is lifted (peak >= 60 % of the recorded palm rise), stays within 30 mm of its in-hand pose from 3 ticks after the grasp stall to the end, never drops more than 2 cm, and is still touched by at least two right hand geoms at the last tick, in the nominal replay and in 10 replays with the target start pose moved by 0.5-2 mm or 0.5 deg (sim_robust_rate = passing fraction). Episodes with sim_success = 0 are kept: the recorded pick succeeded on the real robot, but the replay does not reproduce it.

• sim_success: 158/195
• nominal replay only: 177/195

Each row also has the recording set, base pose, table position and top, and object start poses (x, y, z, yaw in the world frame).