ROBO-SWE

236 executable issue-resolution tasks from 28 robotics software repositories. Each task pairs an original GitHub issue with a repository snapshot before the fix, the developer's reference patch, regression tests, and a container build recipe. The tasks are also distributed in their original Harbor format.

Model leaderboard

All models below were evaluated on the same 236 tasks. The score is the mean percentage of tasks resolved per completed cycle; higher is better.

Model	Harness	Effort	Complete cycles	Resolved (mean)
Opus 5.5	Claude Code	medium	3	56.50%
GPT-6.1 Sol	Codex	max	3	54.10%
GPT-6.1 Sol	Codex	medium	3	52.54%
GPT-6 Astra	Codex	medium	3	51.41%
GLM-5.3	OpenCode	—	1	46.19%
Sonnet 5.5	Claude Code	medium	3	45.76%
Kimi K3.1 Pro	OpenCode	—	2	43.86%

Each complete cycle contains one attempt on every task. The mean is successful attempts divided by all attempts in completed cycles. Partial cycles and interrupted attempts are excluded. Consider the different cycle counts when comparing scores. A dash means no effort setting is recorded. Rows are ordered by the observed mean.

Source: result summaries and all 4,248 complete-cycle results.

Dataset coverage

Domain	Tasks	Repositories
ROS 2 systems	103	17
Simulation and physics engines	63	4
Robotics algorithm libraries	32	5
Robot-learning frameworks	38	2
Total	236	28

This is an evaluation dataset with one test split. There is no separate training or validation split. Reference patches and tests are public for reproducibility; they must be withheld from agents during evaluation.

Load the task table

from datasets import load_dataset

tasks = load_dataset("vharatian/Robo-SWE", split="test")
assert len(tasks) == 236
print(tasks[0]["instance_id"])

Pin the repository commit with revision="<commit>" for a reproducible experiment. data/test.jsonl contains the same records for tools that do not use Parquet.

Field	Meaning
instance_id, repo	Stable task identifier and upstream GitHub repository
domain, domain_code	Domain label and construction-pipeline domain code
pull_request, issue_url	Original pull request and primary issue URLs
base_commit, merge_commit, merged_at	Pinned source revision, fix revision, and merge date
problem_statement	The exact Harbor instruction, including original issue text and workspace guidance
patch, test_patch	Reference production patch and hidden developer-test overlay
FAIL_TO_PASS, PASS_TO_PASS	Measured regression-test identifiers from the task's verifier
statement_kind, file_hint	Original-issue statements; no reference-patch file hints
automated_review	Latest available automated fairness flag, or not_reviewed; not human verification
harbor_archive	Relative path to the complete task archive
source_license_labels	Labels detected from pinned upstream license notices; see full evidence
source_license_evidence, source_commercial_notice	License inventory and additional commercial-notice flag

Some instructions combine multiple linked issues. issue_url names the primary issue; the exact evaluated combined statement is preserved in problem_statement. Do not substitute a freshly fetched issue body, which may have changed.

Download and extract Harbor tasks

Use Python 3.12 or newer for the extraction helper. Download the repository with huggingface_hub.snapshot_download:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="vharatian/Robo-SWE",
    repo_type="dataset",
    local_dir="Robo-SWE",
)

Then extract all tasks, or select one task:

python Robo-SWE/tools/extract_harbor.py --output ./robo-swe-extracted
# Alternatively, into a different fresh directory:
python Robo-SWE/tools/extract_harbor.py --output ./robo-swe-example --task ros2__rosbag2-1812

Each archive extracts as tasks/<instance_id>/ and contains:

instruction.md
task.toml
environment/     Dockerfile, build inputs, network configuration
tests/           verifier, developer-test overlay, measured test identifiers
solution/        developer reference patch and oracle entry point
TASK_MANIFEST.json

The extra manifest records hashes and executable modes. All original task files are preserved byte-for-byte, including the non-ROS source snapshots and dependency inputs. Archives preserve relative symbolic links. One upstream documentation symlink in python-control__python-control-1017 was already broken; it is retained and recorded in provenance/harbor-archives.json rather than silently repaired.

The reference harness version is Harbor 0.24.0. On a Linux machine with Docker, a reference-solution run for the example task is:

harbor run -p ./robo-swe-example/tasks/ros2__rosbag2-1812 -a oracle

This command builds an environment and runs tests. Builds require network access and can take substantial time and disk space; grading runs offline. Task-specific CPU, memory, disk and timeout settings are in task.toml. Limit concurrent ROS jobs to three on a 12-core host: test outcomes can be sensitive to host load. Model-agent runs need an appropriate credential and network arrangement; the archived tasks do not grant general network access during evaluation.

MowgliNext prerequisite images

Seven tasks use two project base images that are not available from a public registry. Their pinned source commits, submodule requirements, build stages and environment settings are included in provenance/project-images.json and each applicable environment/project_env.json.

# List requirements without building:
python Robo-SWE/tools/build_project_images.py --tasks ./robo-swe-extracted/tasks
# Rebuild the two prerequisites, then run the corresponding Harbor tasks:
python Robo-SWE/tools/build_project_images.py --tasks ./robo-swe-extracted/tasks --build

The helper requires Docker Buildx, Git and the GitHub CLI. It uses the same project-image builder as the original pipeline. It downloads pinned public source revisions and submodules; it does not run a model or verifier.

Construction and exclusions

The candidate ledger contains a detailed record for each of the 1,567 candidate pull requests, including saved screening decisions, construction stages, admission outcomes, and rejection/exclusion reasons. Download the detailed JSONL, summary CSV, or validation report. It distinguishes postponed and unprocessed candidates from rejections, retains the original diagnostic granularity, and reconciles 928 queued builds, 489 successfully built environments, and 262 admitted tasks. Post-admission checks removed six tasks at export, nine after harness checks or inspection, and ten after original ROS issue-text screening. One further task became unavailable when its exact nightly dependency disappeared. Thus 262 − 6 − 9 − 10 − 1 = 236.

The unavailable task is google-deepmind__mujoco_warp-1663, whose pinned MuJoCo 3.12.1.dev974703000 wheel could not be recovered. It is excluded uniformly from every published model comparison. The original 237-task frozen membership remains recorded internally; the public evaluated cohort contains 236 tasks. See provenance/exclusions.json and provenance/pinned-versions.json.

During construction, the ROS selector saw issue text, metadata, paths and test names, but no patch. The other-domain selector also saw the developer's test changes, not the production fix. These were screening inputs; evaluated agents received the issue and pre-fix workspace, without the hidden tests or reference fix.

Evaluation records

results/complete-cycles.csv contains 4,248 graded attempts from complete 236-task cycles: three cycles each for Opus, Sonnet, Sol medium, Sol max and Astra; one for GLM; and two for Kimi. results/summary.json gives the corresponding counts and mean resolve rates. Partial cycles and interrupted attempts are omitted from this table. Missing costs remain empty rather than being counted as zero. trial_seconds is the recorded trial duration, not the paper's agent-only runtime.

This dataset release includes task artifacts and the selected result table. It does not include raw agent trajectories or private reviewer records. Those should not be inferred from the presence of the result table or construction summary.

Validation and limitations

All 236 packages are checked against their pinned source folders: complete file inventories, hashes, modes, safe extraction, Harbor parsing, local Docker build inputs, shell syntax and patch syntax. See provenance/harbor-validation.json. No new Docker builds, oracle/no-op runs, model solves or grades were performed for this publication. Prior execution validation does not guarantee that every external dependency or base image remains available for a fresh build today. Some original recipes resolve dependencies dynamically; Docker images themselves are not bundled.

Human fairness review is incomplete. Automated review flags are advisory and do not establish that a task is fair or unfair. Repository and domain coverage is uneven: this is a curated issue-resolution benchmark, not a representative sample of all robotics software or an evaluation of physical robot safety.

Licensing and attribution

This repository uses component-specific licensing; see LICENSE. Original ROBO-SWE contributions are distinguished from upstream material. The complete source-license evidence for all 236 tasks is in provenance/source-licenses.json, with retained notices in licenses/upstream/. Detected labels do not replace the full terms or establish the licensing of every issue, asset or vendored component. MowgliNext includes an additional commercial-use notice; its seven tasks are explicitly flagged. No blanket permissive license is asserted for the bundle.

Credit belongs to the upstream issue reporters, developers and maintainers. Each task links to its source repository, issue and pull request and pins its revisions.

Citation

Authors, in manuscript order: Vahid Haratian, Vahid Ansari, Mohammadhossein Malmir, and Alois Knoll. The BibTeX below uses the standard Surname, Given name format; and separates the four authors.

@misc{haratian2026roboswe,
  title = {ROBO-SWE: An Executable Issue-Resolution Dataset Mined from Robot Software Repositories},
  author = {Haratian, Vahid and Ansari, Vahid and Malmir, Mohammadhossein and Knoll, Alois},
  year = {2026},
  howpublished = {Hugging Face dataset},
  url = {https://huggingface.co/datasets/vharatian/Robo-SWE}
}

A publication DOI is not assigned in this snapshot. Cite the dataset repository and the exact revision used.