Devon018/tilted_insertion_tactile_v5_pcsim4

Tilted Insertion Tactile v5 + 3D-AGOS (pcsim4)

The ready-to-train behaviour-cloning dataset behind the B1-pcsim4 policy (depth_tactile_cnn_unet_agos) of the Insertion-RL repository. It contains:

• the cleaned tactile v5 demonstrations: 49 tilted-insertion assets (asset_700001–asset_700049), and
• the precomputed 3D-AGOS point-cloud v4 sidecar (*.agos_pcsim4.h5) for every trajectory.

You can train straight after extracting the archives. You do not need to rebuild the AGOS sidecars or the plug scans.

Assets	49 (asset_700001 … asset_700049)
Trajectories	5,874 (2,940 normal + 2,934 recovery)
Steps	1,818,774
Files	5,874 trajectory .h5 + 5,874 sidecar .agos_pcsim4.h5
Size	156 GiB to download (49 .tar.zst), 185 GiB extracted

Quick start

# 1. download
hf download Devon018/tilted_insertion_tactile_v5_pcsim4 --repo-type dataset --local-dir ./tactile_v5_pcsim4

# 2. extract the data + plug scans into the repository (needs `zstd`; checks the counts at the end)
bash ./tactile_v5_pcsim4/extract.sh /path/to/Insertion-RL            # add --delete-archives to save disk space

# 3. train B1-pcsim4 (frozen recipe: seed 233, batch 64, 145k steps, EMA) from the repository root
DATA=tactile_full_v5_minmax_agos_pcsim4 OUT_ROOT=outputs/tactile_agos_pcsim4 GPU=0 \
  bash scripts/train_tactile_base_policy.sh depth_tactile_cnn_unet_agos

After extraction the layout is exactly what the data configs expect:

Insertion-RL/datasets/tilted_insertion_tactile_v5/
├── REMOVED_TRAJECTORIES.json
├── asset_700001/
│   ├── disassembly_traj_0.h5
│   ├── disassembly_traj_0.agos_pcsim4.h5
│   └── ...
└── asset_700049/

The first training run builds the lazy cache data/cache/tilted_insertion_tactile_v5_700xxx_depth480x640_force_tactile_f16z (float16, ~161 GiB). The cache is not part of this upload. Force and tactile normalisation statistics are fitted at training time and saved in the checkpoint, so no external statistics file is needed.

Repository files

File	Content
asset_7000XX.tar.zst	One archive per asset. Each holds asset_7000XX/disassembly_traj_N.h5 and .agos_pcsim4.h5.
extract.sh	Extracts every archive into <repo>/datasets/tilted_insertion_tactile_v5/ and checks the file counts.
plug_scans_sim_mv.tar.zst	Multi-view plug scans outputs/agos_plug_scans/sim_mv/asset_<id>.npz (54 assets, 53 MB). Online AGOS needs them for evaluation; extract.sh puts them in place.
manifest.json	Per asset: normal/recovery counts, steps, and every trajectory's id, category and length.
REMOVED_TRAJECTORIES.json	The 6 trajectories removed by the quality audit (SHA-256, measurements, reasons).
configs/	Copies of the data configs used for training (tactile_full_v5_minmax*.yaml, multi_trajectory.yaml). The configs in your repository checkout are authoritative.

Provenance

• asset_700002–asset_700049 come from Hugging Face yinongh/tilted_insertion_tactile_v5, revision 75d37d3333c8229e1bcaa04a27da8e5e8324ef6c. This is the v4 protocol re-collected with the randomized grasp turned on.
• asset_700001 is missing from that upload. It was collected locally with the tracked collector at commit dfeeb2fe.
• Cleaning: a local audit found six trajectory-level physics/force anomalies (pose discontinuities and a sustained 100–1177 N hand-force segment). They are removed from this release; see REMOVED_TRAJECTORIES.json. Four milder anomalies were kept, matching the cleaned v4 selection: three trajectories with all-zero tactile data and one recovery trajectory that never approaches the reference pose.
• AGOS sidecars: built with Insertion-RL commit a363956a, which introduced point-cloud AGOS v4. The algorithm is unchanged at later commits. The build was:

  C="--root datasets/tilted_insertion_tactile_v5 --pipeline pointcloud --plug-scan sim --pc-version v4 --pc-suffix .agos_pcsim4.h5"
  python scripts/build_agos_dataset.py --assets all $C --categories normal
  python scripts/build_agos_dataset.py --assets all $C --make-donor-states --donor-states <donor_dir>
  python scripts/build_agos_dataset.py --assets all $C --categories recovery --warm-start --donor-states <donor_dir>
  

  The build reads the multi-view plug scans outputs/agos_plug_scans/sim_mv/asset_<id>.npz (scripts/scan_plug_multiview.sh). Recovery demonstrations are warm-started from the socket cloud of the best normal approach of the same asset.

Train / holdout split

The split is defined in configs/diffusion_policy/data/tactile_full_v5_minmax.yaml (holdout_ids):

• 5 whole-asset holdouts: 700005, 700013, 700022, 700034, 700046
• 88 explicit trajectory holdouts

In the reference run this gives 5,181 training and 687 usable holdout trajectories. Six trajectories with T = 30 steps are dropped by the min_steps = 30 rule.

File schema

disassembly_traj_N.h5: one episode of length T, played in insertion order.

Key	Shape	dtype	Note
actions	(T, 1, 9)	float32	EEF-frame delta: translation (3) + rot6d (6)
camera3_depth	(T, 1, 480, 640)	float16	wrist depth, negative metres
camera3_rgb	(T, 1, 480, 640, 3)	uint8	wrist RGB
camera3_proj, camera3_vinv	(T, 1, 4, 4)	float32	camera projection / inverse view (world frame)
tactile_depth	(T, 1, 2, 20, 25)	float16	GelSight-Mini, two fingers
tactile_force_field	(T, 1, 2, 20, 25, 3)	float16
contact_force	(T, 1, 8, 3)	float32
fingertip_centered_pos / _quat	(1, T, 3) / (1, T, 4)	float32	env-local, quat xyzw
plug_pos / plug_quat	(1, T, 3) / (1, T, 4)	float32	privileged
init_plug_pos / init_plug_quat	(1, 3) / (1, 4)	float32	seated plug pose

Root attributes: traj_category (normal / recovery), terminal_len, source_run_seed, source_env_id, tactile_sensor_type.

disassembly_traj_N.agos_pcsim4.h5: the AGOS observation for the same T steps.

Key	Shape	dtype	Note
agos	(T, 1, 2, 80, 80)	uint8	two-channel occupancy image (socket, plug), wrist-camera axes
agos_readout	(T, 1, 8)	float32	readout vector (index 3 = valid flag)
agos_fresh	(T, 1)	bool	True on the steps where the estimate was recomputed (every 5 steps)

Attributes: version=agos_sidecar_v1, pipeline=pointcloud, pc_version=v4, plug_scan=sim, every=5, warm_start.

Notes

• This release covers only the data that the tactile_full_v5_minmax_agos_pcsim4 config reads. The other AGOS variants (.agos.h5, .agos_v2.h5, .agos_pcsim{,2,3}.h5, .agos_oracle*.h5, …) and the DAgger recovery data (tilted_insertion_tactile_v5_dagger_rec) are not included.
• The online AGOS used during evaluation (v9 benchmark) reads the plug scans from outputs/agos_plug_scans/sim_mv/. extract.sh restores them from plug_scans_sim_mv.tar.zst, so you do not need to rescan.