chengenwu/isaaclab_tactile

Isaac Lab Tactile GearMesh: demos and policy rollouts

A Franka arm with two GelSight R1.5 fingers (TacSL visuo-tactile sensors) picks the medium gear of Factory's gear-mesh set off the table by its hub and slides it onto a peg of the gear base. The task is a manager-based Isaac Lab environment (Isaac Sim 5.1, Isaac Lab 2.3) in the RoboticArmSim repository (tasks/manager_based/tactile_gear_mesh). The demos come from a scripted screw-motion planner driven through the environment's relative-IK action. The policies are flow-matching policies of the robopolicy library: a 7-level resolution-routed mixture of experts, and an input-mode mixture of experts (all inputs / no RGB cameras / no tactile).

Layout

path	contents
datasets/<name>/	raw demo collections: Isaac Lab recorder HDF5 (data/demo_<i>/{actions, obs/..., states/..., expert/...}) with collection logs
datasets/<name>_500/	the same demos converted to robopolicy's small-files format: tactile_gear_mesh/demo_manifest.json, shards/shard_XXX.hdf5 (actions, robot state, env state), videos_shards/<camera>/shard_XXX.{mp4,json} (cameras table, wrist, tactile_left, tactile_right, 256 x 256)
diagnostics/gear_funnel_diag/	every attempt (successes and failures) of a 20-attempt diagnostic of the converging grasp
rollouts/<run>_eval/	evaluations of one trained policy (see below)
rollouts/logs/	the training and evaluation pipelines' logs

Every raw demo has, per control step (30 Hz): the 7-D relative-IK action plus gripper command, robot state (eef_pos, eef_quat, joint positions and velocities), gear and gear-base poses, the table and wrist camera RGB (256 x 256), and per finger the GelSight tactile RGB (320 x 240), the TacSL force field (20 x 25 x 3) and penetration depths; obs/subtask_terms/{grasp, mesh} mark grasping and success.

Datasets

dataset	task id	demos	how the planner moves
tactile_gear_mesh	TactileGearMesh-Franka-IK-Rel-v0 / -NoFlank-v0	29 (first collections, with previews)	straight descents
tactile_gear_mesh_noflank (+ _500 converted)	TactileGearMesh-Franka-IK-Rel-NoFlank-v0	504	straight descents onto the hub and the middle peg; no neighbouring gears
tactile_gear_mesh_funnel_preview, _funnel2_preview	...-NoFlank-MarkedPeg-v0 (target peg green)	6 each	converging insertion (5-25 mm off at the top, shrinking with height); funnel2 also a converging grasp
tactile_gear_mesh_converge	...-NoFlank-MarkedPeg-v0	471	converging grasp and insertion at the nominal grasp height. Superseded: this grip lets the gear tilt in the hand (see diagnostics/)
tactile_gear_mesh_converge2 (+ _500 converted)	...-NoFlank-MarkedPeg-v0	505	converging grasp and insertion (5-25 mm), fingers 2 mm deeper on the hub (--grasp_depth_mm 2)
tactile_gear_mesh_biggear_preview	...-NoFlank-MarkedPeg-BigGear-v0	10	as converge2, with the gear scaled 1.5x on the unchanged peg of the large gear (2.75 mm of hole play instead of 0.25 mm)
tactile_gear_mesh_biggear_preview_try1	same	2 (1 success)	first BigGear try; superseded (its force field reads zero because TacSL ignored the gear's scale)

Rollouts

One directory per trained policy (resmoe64: 7-level resolution MoE, every image degraded to 1 ... 1/64; modemoe: input-mode MoE), each trained 60 epochs on the dataset of its name (gear: NoFlank, markedpeg: pushed demos with relabelled actions, since deleted, converge2):

file	contents
paired/{ds<f>, mode<m>}.jsonl	success eval: one line per scene (seeds 70000 ...), the same scenes at every level / mode, with milestones
paired/summary.txt	success rate per level / mode
recordings/seeds<s>.h5	recorded rollouts at the policy's first level, one group per scene: per step the policy's inputs, VIG over the other levels (32 shared-noise samples), normalized action chunks, planned fingertip paths, camera matrices, executed action and task state (fingertip, gear and base poses, hub-to-peg offset, gear tilt, grasp / seated signals, gel penetration and force per pad)
flow_rollouts_<lane>.npz	flow-matching loss (512 draws) per recorded step
demo_frames_first50_stride4.h5, vig_demos.npz, flow_demos.npz	every 4th frame of the first 50 training demos and the scores on them (thresholds)
stage_analysis.json	per-stage labels and statistics of the recorded rollouts

Results of the converge2 policies (paired, 50 scenes per level / mode): input-mode MoE 28 % with all inputs, 0 % without RGB, 8 % without tactile; resolution MoE 18 / 8 / 16 / 20 / 14 / 14 / 0 % from full resolution to 1/64. Most failures lower the gear onto the peg a few millimetres off its axis and then press on the peg top instead of re-centring.