--- title: robodojo-tactile-memory canonical_url: "https://www.modelscope.cn/datasets/genalyu/robodojo-tactile-memory" md_url: "https://www.modelscope.cn/datasets/genalyu/robodojo-tactile-memory.md" repository: genalyu/robodojo-tactile-memory last_updated: 2026-10-30 license: apache-2.0 storage_size: "142 GB" downloads: 12376 stars: 0 --- # robodojo-tactile-memory > robodojo-tactile-memory - genalyu 在 ModelScope 开源的数据集。RoboDojo Memory six-task dual-Franka optical tactile simulation port. Synchronized RGB, gel tactile images, state and actions. Incremental collection; release completeness and limitations are recorded in… genalyu/robodojo-tactile-memory 是 ModelScope 魔搭社区上的数据集，存储大小 142 GB，采用 apache-2.0 许可。 - **Repository**: genalyu/robodojo-tactile-memory - **License**: apache-2.0 - **Storage size**: 142 GB - **Downloads**: 12376 - **Stars**: 0 - **Last updated**: 2026-10-30 Source: https://www.modelscope.cn/datasets/genalyu/robodojo-tactile-memory --- # RoboDojo-Tactile Memory 同步视觉与凝胶光学触觉示范数据，覆盖 Memory 六任务：cover_blocks、swap_blocks、swap_T、press_by_number、imitate_sorting_sequence、match_and_pick_from_conveyor。 **当前为持续上传中的增量版本，尚非完整基准发布。** 完整目标为每任务 100 条训练、20 条独立验证，共 720 条成功示范。实际已验证上传数量见 `manifests/index.json` 的 `counts_by_task`，仅 `release_complete: true` 表示数量齐备。评测种子独立冻结，不按专家成功筛选。 ## Download and reuse Install the official client into your own Python environment: ```bash pip install modelscope-hub==0.5.1 ``` Download and extract `code/robodojo-tactile-memory-code-a3ede5e.tar.gz` from this repository. Its SHA256 is `615cfc0b5c662fa55607805141061ebb14cd66b25090059b583dd68d6328f167`. ```bash python scripts/tactile/restore_modelscope_archive.py --out /data/robodojo-tactile/neosim ``` Use `--limit 1` for a small recovery test. Add `--require-complete-release` when a full 720-episode release is required; partial inventories are rejected. The index pins each HDF5 to an immutable commit and SHA256. The restore command checks full downloaded bytes, not just HTTP success. Keep `restored_inventory.json` with your training run. Set `MEMORY_DATA_ROOT=/data/robodojo-tactile/neosim` and follow `docs/N0_MEMORY_REPRODUCE.md` inside the code archive. Train and validation are prepared independently; normalization uses training only. Dependencies and base weights are installed separately under their own licenses. Frozen collection sources are under `sources/`; procedural scene layouts and source identities are also embedded in each HDF5. ## Data contract Each successful complete HDF5 preserves 30 Hz synchronized three RGB cameras, four clean gel tactile RGB streams, per-arm robot state, simulation clock, and auxiliary sensor records. Actions for n0twam are the next observed end-effector state, in each robot base frame; the final row has no next-state label and is dropped. Imitation controls A only while B demonstrates. Use the provided adapter to enforce this protocol. NeoSim JPEG storage has a native RGB/OpenCV convention: decode with the adapter, without an extra BGR/RGB swap. Clean tactile RGB is the policy input. Raw marker coordinates are uncalibrated auxiliary data and are not claimed as valid shear/marker tracking. Raw UIPC contact gradients are not calibrated forces in newtons. ## Scope and validation This is a dual-Franka port of the Memory task semantics, with NeoSim/UniVTAC TacEx/Taxim gel sensing. It differs from native RoboDojo ARX embodiment and some fixtures; scores must be reported as port-specific. See `docs/tactile/NEOSIM_MEMORY_PROTOCOL.md` for rule differences. Before a server copy is removed, each episode is audited, uploaded, downloaded from its committed version, SHA256 checked, and recorded under `manifests/episodes/`. The first downloaded trajectory was also read through the n0twam adapter: 377 state/action rows, three 270x480 RGB streams, and four 240x320 tactile streams. This storage validation does not imply policy convergence. Short training and closed-loop integration have been exercised; a completed formal six-task learned-policy score is not yet available. The dataset repository uses Apache-2.0. The supplied project code retains its own LICENSE and upstream notices. Isaac Sim, cuRobo, n0twam code/weights, and third-party assets retain their respective terms; the dataset license does not relicense them. The updated code archive includes a shared-normalization six-task pool builder and per-task controlled-arm loss masks. See `docs/N0_MEMORY_REPRODUCE.md`. Collection and uploads remain in progress; only the inventory release gate establishes completeness. Code 509c5ce passes 87 checks against the pinned public n0twam core. Single-arm imitation masks converted B-arm padding before loss calculation; serving uses normalization-bound neutral padding. See the included action-contract evidence. Code 7e4f36d also pins the publicly available exact UniVTAC/TacEx source and resource recipe, and includes a complete-task streaming preparation command. Six-task optical sample QA is available under [qa/optical_v01](https://www.modelscope.cn/datasets/genalyu/robodojo-tactile-memory/files?path=qa%2Foptical_v01). It compares every clean gel RGB frame with the initial frame in one pinned successful trajectory per task. Active pads changed in all six samples, with static support pads unchanged. This sample check does not replace the full release audit. The code archive includes the measurement command and exact source identities. Incremental expert attempt snapshots are under manifests/collection/: failures, unresolved attempts and storage-unavailable successful episodes are accounted for separately. The supplied snapshot command and its checks are included in the updated code archive; 89 project checks pass against the pinned public n0twam core. Collection scratch: NeoSim temporarily writes uncompressed per-frame pickles before HDF5 completion. A measured imitation trajectory needed about 9 GB of scratch. The included reproduction guide explains dedicated RAM frame caches; the recovered worker passed three native diagnostics and an audited production sample with verified cloud readback and pruning. This does not establish full-release completeness or a learned-policy score. ## Raw dataset release status — 2026-10-30 All 720 raw trajectories are published: each Memory task has 100 training and 20 validation episodes, totaling 151,002,387,112 bytes. Each file was downloaded at its immutable revision and fully SHA256 verified before the server original was removed. The public inventory was anonymously read back; it has 720 unique task/layout pairs and SHA256 cb126dce42d0920b0d697e7f5b04dc2e1ca0471dbe967f24dec29a4e812995df. See qa/raw_release_v01/verified.json and manifests/index.json. Derived VAE/DINO encoding is still in progress and is not included in this raw completeness claim. A learned-policy benchmark score is not yet available. The latest code includes batched upload/readback, native HDF5 decoding, complete-release restore/audit commands and the independent train/validation preparation entry points. Dependencies and upstream models retain their own licenses.