--- pretty_name: ARISE Real-to-Sim Scenes and Demonstrations language: - en - zh tags: - robotics - real-to-sim - simulation - robot-manipulation - video size_categories: - n<1K configs: - config_name: demonstrations data_files: - split: data path: metadata/episodes.jsonl - config_name: scenes data_files: - split: data path: metadata/scenes.jsonl --- # ARISE Real2Sim Four reconstructed robot-manipulation scenes and their **40 original demonstrations**: ten source layouts per task. This is one release repository, with exactly one scene version for each task. Coin insertion is excluded. | Task folder | Demonstrations | Selected scene | |---|---:|---| | `pick_the_moving_cube` | 10 | moving-strict20 | | `press_rotating_clock` | 10 | clock-calibration27 | | `put_bottles_into_the_bin` | 10 | rim14-reflective-20261116 | | `stack_blocks_medium` | 10 | stack-control-parsimony-long120 | These are the current retained task-specific versions selected for this release, not a claim that one score ranks reconstruction quality across tasks. The requested bottle version has a 14 mm rolled lip, reflective bottle materials and a broadened shared light. Its new full recorded-command replay contains physical deviations and **does not pass the complete task predicate**. Its geometry/material/light readback and result are included in `replay-evidence.json`. ## Contents ```text scenes//scene.yaml one base scene per task scenes//layouts/*.yaml ten source-layout overlays, not ten method versions demonstrations/// original MP4, NPY, metadata and recording flags demonstrations//layouts/ source layout images assets/ meshes, textures, robot resources and Gaussian fields runtime/ matched source snapshots for the scene schema/loader previews/ retained scene views and the new bottle replay metadata/ episode mapping, versions, resources and checksums tools/prepare_scene.py CPU preparation; no simulation starts automatically ``` Scene resource references use `bundle://` relative to this repository. Local interpreter, CUDA, renderer and cache locations use `machine://` placeholders. Robot USD layers and textures are localized; the installed Isaac material modules, such as OmniPBR, remain simulator dependencies. No Isaac installation, pretrained model checkpoint, private policy, token or machine configuration is included. ## Original demonstrations All 40 episode directories retain the source file bytes. There are 140 original camera videos: Top, Left and Right for all episodes, plus the two recorded Gemini views in the ten stacking episodes. Actions, follower states, controller/state timestamps, camera timestamps and metadata remain in their original units and formats. No action or clock has been resampled. Use `metadata/demonstrations.json` to map task, layout and episode IDs. Camera timestamps and controller/state clocks must be interpreted separately; encoded video seconds are not exact exposure times. The public `data` index is an inventory, **not a newly assigned train/test split**. Preserve an experiment's original split and prior-exposure records when using these released demonstrations. ## Download ```python from huggingface_hub import snapshot_download snapshot_download("YiyangJin/ARISE-Real2Sim", repo_type="dataset", local_dir="ARISE-Real2Sim") ``` Download a task subset by selecting its `scenes/` and `demonstrations/` folders together with `assets/`, `runtime/`, `tools/` and the root metadata. Asset files shared across tasks are stored once. ## Prepare a scene Use Python 3.11 with the existing ARISE dependencies. The native integration requires an independently installed compatible Isaac Sim 5.1 / RoboDojo environment. Gaussian rasterization additionally requires its explicitly configured, isolated gsplat environment. No model or dependency is downloaded by the preparation helper. ```sh python tools/prepare_scene.py --task put_bottles_into_the_bin --layout four_08 --output /tmp/arise-bottles # For the raster-Gaussian tasks, first edit a local copy of machine-config.example.yaml. python tools/prepare_scene.py --task pick_the_moving_cube --layout 01 --machine-config /path/to/local-machine.yaml --output /tmp/arise-moving # Optionally compile a new owned native runtime from an existing RoboDojo checkout. python tools/prepare_scene.py --task put_bottles_into_the_bin --layout four_08 --robodojo /path/to/RoboDojo --output /tmp/arise-bottles-runtime ``` The helper retains scene parameters, applies the chosen source-layout overlay, resolves resources and validates with the **matched source snapshot**. It does not launch simulation, a policy or a GPU experiment. Layout-specific observations and initial conditions are preserved; a newly prepared layout is not a new native validation result. ## Validation scope The release checks exact source demonstration hashes, four scene versions, forty unique episode mappings, resource closure, robot physical-attribute preservation during relocation, and matched schema/CPU preparation. The bottle preview is a fresh complete physical command replay; the other scene previews are retained views from the selected source versions. This publication does not add policy trials or certify real/sim success-rate equivalence. Estimated geometry, dynamics, materials and unobserved surfaces retain their recorded provenance. ## Source and resource terms The ARISE harness source is at [SII-LiuLab/ARISE](https://github.com/SII-LiuLab/ARISE). Matched source snapshots are retained here because scene contracts and native execution support evolved between experiments. Native RoboDojo files retain their accompanying MIT license. Robot resource provenance identifies the upstream YAM/i2rt inputs. Isaac/Omniverse and other external tools retain their own terms and are not redistributed as installed software. No new dataset-wide license has been selected for this upload. The original recordings and derived scene resources are released at the owner's request; upstream notices and resource-specific terms remain applicable. Release: **2026-11-16**. `metadata/SHA256SUMS.txt` covers every release file except itself.