magictactile-sharpa-aligned-handoff

Hand-over package of the ShARPA (SharpaWave hb1 / ha4) aligned-teacher work done on nebula (session sharpa-soft), 2026-10-10 .. 2026-10-12. The machine that produced it cannot be reached by ssh from the new one, so everything travels through this dataset repo (private).

Start with docs/handoff/HANDOFF.md (master entry / index). Then: current state 01_CURRENT_STATE.md, problems 02_PROBLEMS.md, data to collect 03_DATA_TO_COLLECT.md, training 04_TRAINING.md, environment 05_ENVIRONMENT.md, runbook (how to resume the replays) 06_RUNBOOK.md, migration 07_MIGRATION.md, lessons 08_LESSONS.md, repo map 09_REPO_MAP.md; notes/ and env/ hold the detailed notes and environment files. docs/handoff/MIGRATION_STATUS.md is regenerated by the uploader and tells what has been uploaded so far (STATUS_AUTOGEN.md is the replay status, also regenerated automatically).

folder	content
code/	MagicTactile_main.bundle = full history of local main (every sharpa/ branch merged into it) and lfs_objects_sharpa.tar = the Git-LFS objects of data/sharpa/** (the ShARPA S/E models), so the repo can be restored without GitHub (tools/restore_repo.sh)*; also a git bundle of every local sharpa/* branch (MagicTactile-sharpa-soft = replay/teacher/student pipeline on branch sharpa/soft; MagicTactile-sharpa = Isaac collection + paired pipeline on sharpa/collect), filtered git archive of both working trees (no data/ tree, no vendor manual), patch series of the local unpushed commits on top of origin/main, untracked files, loose scripts that live outside git (code/outside_git/), COMMITS.txt
checkpoints/	new aligned S / E / P / hybrids (r2, r1b), the old round-10 warm starts and deployed bundles, soft-object bundles; INDEX.md has sha256 and metrics
assets/	aligned geometry / profile / camera bundles for hb1 and ha4, sysid tables, press series, frozen selections and episode lists (selections_relocatable/ with a {DATA_ROOT} placeholder), real-range probe results, query_reference.npz, the public SDK (assets/sdk/: ONNX models, embedded maps, sample frames)
data/inputs/inputs_NNNN.tar	REPLAY INPUTS of every episode of the four selections (pilot, v8, deep, rest; 2901 episodes): the old Isaac collection (<campaign>/<shard>/collection/{collection.json, <env>/*}) and the camera-independent Isaac query features (<campaign>/<shard>/paired/<env>/query/*); extract into the new sharpa_data/ root, then align_batch.py can (re)play any episode
data/caches/<name>/<name>_NNNN.tar	optional, only if uploaded: the FINAL training caches (cache_v2column, cache_v2column_far), tar paths start with <name>/; extract into the data/<TAG>/ folder (see tools/uploader/pack_caches.py)
data/isaac/<campaign>_NNNN.tar, <campaign>_paired_NNNN.tar, <campaign>_extras.tar	Isaac-only collections (v10_env: 124 shards + 2 pilots; cannot be recreated without Isaac Lab): <campaign>/<shard>/{collect.log, collection/**}, the query features computed so far (paired_*/<shard>/<env>/query), recipes / csv / json / logs / selection files; extract into the new sharpa_data/ root
data/soft/<group>_NNNN.tar, evals.tar, README.md	soft-object RAW data (MPM / rope / hand / patch worlds with their IPC teacher outputs, small checkpoints, exports; ~57 GB), paths relative to sharpa_data/soft/; the soft pipeline needs the Newton venv (docs/handoff/env/requirements_newton_venv.txt); see data/soft/README.md
data/replays/<root>/<root>_NNNN.tar	tar shards of the FINISHED aligned IPC replays (states, forces, labels, teacher images, query features, collections), uncompressed tar (the npz files are already compressed), + *.manifest.json per shard
tools/	download_all.py, unpack.sh, verify.py, rewrite_paths.sh, make_selection_local.py, and uploader/ (the scripts that produced this repo)
docs/	hand-over documents, the alignment report (sharpa_align), progress log

Download

from huggingface_hub import snapshot_download
snapshot_download("luhr/magictactile-sharpa-aligned-handoff", repo_type="dataset", local_dir="handoff_repo", max_workers=8)   # resumable

or python tools/download_all.py --dest handoff_repo [--only docs code assets checkpoints data/inputs data/replays/hb1_v8] after fetching tools/ alone. Log in first (hf auth login or HF_TOKEN).

Unpack to the original layout

bash tools/unpack.sh handoff_repo  /new/root/tactilelab      # inputs + finished replays + code + assets, in the layout all scripts expect
python tools/verify.py handoff_repo                           # sha256 of every shard against its manifest
bash tools/rewrite_paths.sh /new/root/tactilelab /new/root/tactilelab/code/*   # the scripts hard-code /wang_ssd/vcj9002/luhr/tactilelab
python tools/make_selection_local.py assets/selections_relocatable/selection_v8.json out.json --data-root /new/root/tactilelab/sharpa_data

Restore the git repository (no GitHub needed)

bash tools/restore_repo.sh handoff_repo /new/root/tactilelab/code        # -> /new/root/tactilelab/code/MagicTactile at `main`, models of data/sharpa/** restored and checked (sha256 = LFS oid)
# by hand:
GIT_LFS_SKIP_SMUDGE=1 git clone -b main handoff_repo/code/MagicTactile_main.bundle MagicTactile      # (a bundle has no LFS server, hence the skip)
tar -xf handoff_repo/code/lfs_objects_sharpa.tar -C MagicTactile/.git                                 # lfs/objects/aa/bb/<oid>
cd MagicTactile && git lfs checkout 'data/sharpa/**'                                                  # real files from the local object store
git remote rename origin bundle && git remote add origin <your remote>

Only the data/sharpa/** LFS objects are shipped (the DIGIT / GelSight data under data/gelsight/** is not, and git lfs fsck will complain about it; ignore). The per-branch bundle code/MagicTactile_sharpa-branches_*.bundle restores the local sharpa/* branches: git fetch <bundle> 'refs/heads/*:refs/remotes/bundle/*'.

Licences, what is NOT here

• The SharpaWave SDK files in assets/sdk/ (ONNX models, embedded point/normal maps, sample frames, manifest) are public and Apache-2.0 ("Copyright 2025 Sharpa Group", release https://github.com/sharpa-robotics/sharpa-wave-sdk/releases/tag/v5.0.10); LICENSE.sharpa-tacmap.txt and NOTICE.sharpa-tacmap.txt at the repo root must travel with them (Apache 4(d)). The SDK can also be downloaded from that GitHub release; the .deb packages and libsharpa*.so are NOT in this repo.
• real robot data (RoCo / T-Rex tiles) and the vendor user manual are not included; teacher/full_states.npz (57 MB per episode, audit only), engine / warp caches and *.workdir are not included.
• ha4 episodes: regenerate them from the hb1 states with scripts/sharpa_soft/align/convert_ha4_all.sh <hb1 root> <ha4 root>.
• training caches (cache_v2column, far/): rebuilt from the episode shards with align_dataset.py + pipeline/prepare_aligned.sh (about 1-2 h of CPU).
• the camera bundles in assets/ contain one reference frame each (from a real hb1 session) and per-pixel position maps derived from the SDK.