GeorgiaTech/ik-modulation-zarr

IK-Modulation — Segment Zarr

Segmented, contact-cycle Zarr datasets for force-conditioned IK-target modulation on a Unitree G1 humanoid. Each task pairs teleop cycles (which carry a real IK "press") with human-demonstration cycles; a transfer rule turns teleop IK offsets into per-hand XY press targets used to train the modulator.

Raw source bags: GeorgiaTech/Lidar-Tactile.

bucket/ — bucket task (2026-05)

Contact label space 1=left, 2=right, 3=both; training cycle 2→3→1.

bucket/
  teleop/teleop_old_ik_231_sequences_bucket_dur4to7p5_closure0p12.zarr   # final teleop (39 cycles)
  human/human_bucket_force_231_sequences_bucket_closure0p20.zarr          # final human  (~24–29 cycles)
  intermediate/
    teleop_old_ik_231_sequences_ikbar0p020_7p5_bucket.zarr   # raw IK-error 2-3-1 cycles (97)
    teleop_old_ik_231_sequences_bucket_dur4to7p5.zarr        # + duration 4–7.5 s filter (63)
    human_bucket_force_segments_all_bucket.zarr              # raw force segments (188)
    human_bucket_force_231_sequences_bucket.zarr             # assembled 2-3-1 (28, pre-closure)
    human_bucket_force_23132_sequences_bucket.zarr           # longer closing 2-3-1-3-2 cycles (sanity)

• Teleop segmented by IK-tracking-error; human segmented by tactile force (≥1 N, sensor1=left). The two final zarrs are what training consumes.
• Zarr layout: data/state (pose + tactile + force), data/action (teleop IK target), data/segment_label (1/2/3), data/force_contact, meta/episode_ends.

import zarr
z = zarr.open("bucket/human/human_bucket_force_231_sequences_bucket_closure0p20.zarr")
print(list(z["data"].keys()), len(z["meta/episode_ends"]))

Full pipeline (segmentation, IK transfer rule, training) is documented in ik_modulation/configs/bucket_pipeline/BUCKET_PIPELINE.md (in the diffusion_humanoid_train repo).

Other tasks in this repo

Sibling datasets from the same IK-modulation line of work also live here, e.g. 0517_human_yogaball/, 0510_yoga_teleop_processed/, 0512_yogaball_human_processed/, and other teleop captures. The bucket/ tree above is the 2026-05 bucket task.