README.md

---
license: cc-by-nc-4.0
task_categories:
- robotics
tags:
- video
- robotics
- maniskill
- simpler-env
- simulator-oracle
pretty_name: SIM Oracle Working Interventions v1
configs:
- config_name: default
  data_files:
  - split: train
    path:
    - data/metadata.jsonl
    - "data/train/**/*.mp4"
  drop_labels: true
---

# SIM Oracle Working Interventions v1

This is the **simulator-oracle** edition of the action-patching pipeline.  It
uses privileged simulator metrics and is not deployment eligible.  It does not
use metric estimators.  The repository also describes the non-oracle,
metric-estimator path; these videos must not be represented as evidence that
those estimators recovered the same signals.

The dataset contains 132 of 132 exposed paired
episodes (complete).  Every row
has a full baseline MP4, full intervention MP4, and full labelled side-by-side
MP4.  Videos start at state frame 0 and end at terminal state frame T.  The two
arm MP4s decode byte-identically through the state frame immediately preceding
the first modified action, inclusive.  Natural versus canonicalized prefixes
and the episode-2339-style action-splice counterfactual are disclosed per row.

These MP4s are deterministic **simulator action-trace re-renders**, not footage
recorded by the original evaluations.  The source runs did not save RGB or
video.  This builder replayed their sealed float32 executed-action traces, with
no policy inference, from the exact source initial state and required the same
final state hash and terminal outcome before admitting each natural arm.

## Scope and outcomes

This is the complete exposed cohort from four accepted held-out validation
runs.  It includes every exposed correction, regression, unchanged success,
and unchanged failure—there is no outcome cherry-picking.  Tuning episodes,
rejected candidates, and unexposed validation episodes are intentionally out of
scope.

- Corrections: **69**
- Regressions: **2**
- Exposed unchanged successes: **41**
- Exposed unchanged failures: **20**

### Tasks

| Task key | Environment | Pairs |
| --- | --- | ---: |
| `cube_stack` | `StackGreenCubeOnYellowCubeBakedTexInScene-v1` | 43 |
| `spoon_cloth` | `PutSpoonOnTableClothInScene-v1` | 89 |

### Accepted patch lineages

| Lineage | Search/authorship protocol | Pairs | Corrections | Regressions | Patch SHA-256 |
| --- | --- | ---: | ---: | ---: | --- |
| `cube_any_grasp_gentle_validation_v1` | fixed candidate search + automatic paired selection | 22 | 13 | 0 | `b64c2a27be859bb88fa0c15fb4a69b42e3eb0aefe4c5b9462161972075725d6c` |
| `spoon_hybrid_regrasp_servo_validation_v1` | fixed candidate search + automatic paired selection | 42 | 22 | 0 | `ac482f06a158d5ac5c5627e8721de9c63db37dfff9db8372161519defd798dd0` |
| `spoon_agentic_premature_release_validation_v1` | autonomous catalog-guided failure loop | 47 | 23 | 1 | `0bccbed2ccf63ee2aa2a96689c2919b606a8a03453906fc7b9ebef1b3197675a` |
| `cube_open_autonomous_validation_v1` | prior-free category + declarative-program synthesis | 21 | 11 | 1 | `112a369517957439dcc7bc922286baf475c39222c9473fe9869be78d226e9699` |

The two fixed-search lineages are useful reviewed-library baselines.  The Spoon
agentic lineage autonomously chose from that catalog.  The open Cube lineage
used neither known-category nor known-patch priors: it synthesized a new
failure predicate and bounded declarative action program, then passed the same
common-inventory tuning and single disjoint-validation gate.

## Videos and causal-prefix convention

Each arm is H.264/yuv420p at 640x480,
10 fps, with 61 frames for 60 actions.  The comparison is
1280x720; baseline
is always left and intervention right.  Frame `t` is the state before action
`t`, so the first edited action executes after the identically displayed frame
`t`; physical divergence can first appear at frame `t+1`.

Of the 132 published pairs, 129 replayed with a naturally exact
RGB prefix.  3 legacy pairs had small pre-edit visual
drift and use a transparently recorded baseline-prefix presentation splice so
the published arms obey the exact-prefix convention.  One of those—fixed-Spoon
episode 2339—also had historical arm-action divergence before its recorded edit
and therefore uses a separately simulated canonical counterfactual: baseline
actions before step 27 and the sealed intervention suffix thereafter.  It is
not represented as a naturally clean source pair.  Per-row fields preserve the
natural-prefix status, canonical splice step, published trace kind, state/action
hashes, and source versus published outcome.  Total canonical action splices:
1.

## Dataset layout

`data/metadata.jsonl` is Hugging Face VideoFolder metadata.  It exposes exactly
three explicitly named video columns: `comparison_file_name` is the labelled
comparison, while `baseline_file_name` and `intervention_file_name` are the two
individual arms.
All paths are POSIX-relative to `data/metadata.jsonl`.

Each row also includes a stable `pair_id`, task/environment identity, episode
and batch placement, patch/formula digests, first intervention step, source and
published effect labels, and causal-prefix mode.  Full per-pair provenance is
in `episodes.jsonl`, `interventions.jsonl`, and each episode's `metadata.json`.

Formulas are mechanically reconstructed from canonical patch config or a
compiled patch-program AST.  Full formula JSON/text is in `formulas/`; the
side-by-side MP4 carries the concise exact action/trigger form and its formula
digest.  For historical legacy runs, the oracle runtime source was not part of
the source manifest; the packaged current runtime is labelled reconstruction
provenance, while sealed executed action traces remain authoritative.

## License and attribution

Dataset metadata and paired derivative videos are released under
[CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/).  Simulation,
robot assets, environments, and policy components retain their upstream
licenses.  Please attribute ManiSkill, SIMPLER/SimplerEnv, Octo, and the
underlying Bridge/WidowX task assets as applicable.  This dataset is
non-commercial and simulator-only.

See `LICENSE_AND_ATTRIBUTION.md` for component-specific attribution and pinned
upstream revisions.  This release does not redistribute or relicense policy
weights or upstream software.

## Integrity

`validation_report.json` records cohort/video/prefix/size gates.
`SHA256SUMS` seals every material file except itself.  The configured hard size
ceiling is 4831838208 bytes (below 5 GiB).

## Limitations

This release covers one frozen policy, two manipulation tasks, and privileged
simulator-oracle signals.  It demonstrates intervention behavior in simulation;
it does not demonstrate estimator recovery, real-robot safety, or deployment
eligibility.  Every new task or invented program requires its own fresh paired
validation.