AHOPPA: initial 64-episode records

Agentic Humanoid Operation through Perception, Planning, and Action.

Kingston Kuan¹ · Samuel J Chew² · Brandon Ong¹

¹ Hebbian Robotics · ² Carnegie Mellon University

This project is not endorsed by Carnegie Mellon University.

Code and paper · Project page · Task and scoring

This dataset contains the retained records for exactly the 64 scheduled starts reported in AHOPPA's initial comparison: four models, two reasoning settings, eight drawer conditions and one attempt per cell. It includes unsuccessful episodes and all six API-failed starts. It contains no pilot, development, controller-check or follow-up experiment records.

The task uses a simulated Unitree G1 to inspect four cabinet drawers, push open drawers closed by hand, leave closed drawers alone and return home before declaring completion. Models operate through visual feedback and numerical head/wrist and locomotion commands to a fixed SONIC controller. These results were obtained in the last few weeks following frontier model releases. No hardware experiments are included.

Downloads

Path	Contents
data/episodes.csv	The public ledger, one row per scheduled start
data/summary.json	Public aggregate and episode-level results
data/provenance.json	Provenance of the public results
episode-summaries.tar.zst	Small JSON records for all episodes, including outcomes, timing, configurations and source manifests where recorded
episodes/<episode>.tar.zst	One raw-record archive per episode
metadata/archives.csv	Episode archive paths, compressed sizes and SHA-256 hashes
metadata/files.jsonl	Original SHA-256 and size of every included regular episode file
metadata/cohort-*.json	The sixteen original frozen comparison manifests
metadata/schedule.json	Frozen condition schedule, with the host workspace path removed
metadata/launches.jsonl	Launch ledger, with node addresses removed
metadata/release.json	Release scope and frozen source identity
metadata/omissions.json	Explicit list of omitted files and metadata changes
metadata/symlinks.json	Original internal symbolic links retained in the archives
SHA256SUMS	Checksums for the release files

The archives use Zstandard compression. Install zstd before extracting them. Download only the episodes needed for an analysis. For example:

hf download hebbianrobotics/AHOPPA --type dataset \
  --include 'data/*' --include 'metadata/*' --include 'SHA256SUMS' \
  --include 'episode-summaries.tar.zst' --local-dir ahoppa-records
cd ahoppa-records
tar --zstd -xf episode-summaries.tar.zst

To download and extract one complete episode:

hf download hebbianrobotics/AHOPPA --type dataset \
  --include 'episodes/storage-cmp3-astra-low-open-a-d-r1.tar.zst' \
  --local-dir .
tar --zstd -xf episodes/storage-cmp3-astra-low-open-a-d-r1.tar.zst

For an exact paper reference, pin the dataset commit or the initial-64-v1 tag using hf download --revision initial-64-v1. The checksum file covers all release payload files other than itself. If the entire release is downloaded, run sha256sum -c SHA256SUMS. For a partial download, verify the selected archive against its row in metadata/archives.csv.

Each archive extracts to episodes/<episode>/. Keep that directory structure. To use the frozen cohort manifests, copy metadata/cohort-*.json to the extraction root; their episodes/... paths will then resolve. A full cohort requires all four corresponding episode archives. Extracting summaries alone does not provide the full evidence required by the cohort validator.

Episode contents

• episode.json: terminal execution status, independent task outcome and timing classification.
• inputs/: frozen scene, facility map, controller constants, project settings and lockfiles, subject to the robot-description omission below.
• source/ and source-manifest.json: the captured runtime source and hashes.
• physics/records.mcap: indexed physical trace, model observations, command ledger and terminal evidence.
• shared-physics/: owner timing, controller logs and native numerical state and command logs.
• operator/: exact recorded request payloads, prompts, input images and arrays, provider responses, tool rounds and timing, where execution reached them.

Some records are stored inside MCAP attachments rather than as loose files. Use the repository's recording readers to access them. Failed starts may lack records that startup never produced. The dataset preserves that absence rather than inventing outcomes. Recorded provider reasoning summaries are provider outputs, not raw chains of thought or causal explanations of model decisions.

Historical webolt identifiers and local file paths are retained inside episode records to preserve their original bytes and hashes. Those paths describe the original runtime and are not paths to use on a new machine. Internal symbolic links stay within their episode directory. The artifacts contain JSON, MCAP, PNG, NPY, CSV and source files; interpreting source is separate from running it.

Evidence integrity and omissions

Every included regular episode file is checked against its original SHA-256 while packaging. Model interaction records and MCAP recordings are distributed without content edits. The original local evidence is unchanged.

The release omits:

• Generated Python bytecode and __pycache__ directories.
• Upstream robot-description .urdf files, including the generated articulated description. These remain identified by the original manifests and runtime configuration but are not relicensed by this dataset.
• Infrastructure launch/deployment YAML and unrelated local experiment folders.
• Generated offline replays, available separately through the public project's hash-pinned media release. Recorded model images remain in the raw archives.

The schedule's workspace path and launch ledger's node addresses are removed in the separate public metadata copies. metadata/omissions.json lists every omitted episode file and these metadata transformations. Other historical local paths in episode records remain unchanged.

Full validation of retained input hashes requires restoring the original inputs/robot.urdf from the separately obtained upstream robot assets. Its expected SHA-256 is 6e107391d015ab27438026bb953b21ef8fa6a02305e65393985908941d7fd98d. Do not substitute a newer robot description. Robot asset bundles, meshes, controller weights, installed simulator environments and API credentials are not supplied. See the public repository's reproduction guide.

For analysis of the original experiment, use the captured source and lockfiles, or the hash-pinned frozen-runtime source asset linked in metadata/release.json. The maintained runtime can differ from the recorded study. No simulator, GPU or model API access is needed simply to read retained records; new experiments require those external dependencies.

Interpretation and limitations

Success requires the model's completion declaration and the independent physical criterion. Infrastructure completion, physical task outcome and timing validity are distinct quantities. The reported success denominators retain all seven scheduled task attempts per model and effort, including API errors; the all-closed condition is a separate control.

Only 17 of 58 completed episodes met the recorded 100 ms physics-lag limit. Timing-invalid episodes remain included and labelled. The experiment uses one fixed layout and one attempt per model/effort/condition. Frames, model decisions, controller ticks and drawers within an episode are not independent trials. The records support analysis of these observed runs; repeated experiments are needed to estimate reliability. Simulation does not establish hardware performance or safety.

Cost and timing

The paper and project page report estimated API cost, full episode duration and model-decision latency for all seven scheduled task attempts and for successful attempts only. The all-closed control is separate. A dash means no successful attempt was observed. The website adds an asterisk to that dash; the paper instead uses a cost asterisk for unresolved usage. Costs use recorded and provider-reconciled token usage at published standard rates, rather than invoice amounts. Astra high was cheaper in these runs because cached reads covered about 50% of its input, compared with 29% at low effort.

The cost and timing method provides the definitions, pricing policy, recovery evidence and reproduction commands. Provider reconciliation is published separately in the repository; these raw archives preserve the original responses, including missing responses from interrupted calls. Eleven Gemini calls remain unpriced. Full episode duration includes inference while physics runs and ends at recorded termination. Decision latency is the median of per-attempt medians for completed decisions. The paper reports all task attempts; the page also shows successful attempts separately. Physics lag is a separate simulator timing quantity. Cost per success in the paper includes spending on failed task attempts and divides total task spending by the success count, unlike mean cost conditional on success.

Attribution and licenses

Original results, numerical records, prompts and author-created presentation content are released under CC BY 4.0. Original project code retains the MIT license. Third-party assets and their depictions retain applicable upstream terms; no rights to upstream weights, geometry or software are granted by the dataset license. See LICENSE.md, CODE-LICENSE.txt and THIRD_PARTY_NOTICES.md.

Credit AHOPPA, Kingston Kuan, Samuel J Chew and Brandon Ong. Cite the paper linked from the public repository and identify the dataset revision or initial-64-v1 release when reporting analyses.