🧪 Bench2Dex Leaderboard Logs

Community evaluation logs for the Bench2Dex leaderboard. Submit your policy's README, results JSON, and organization logo separately to the GitHub leaderboard repository.

📦 Submit your policy logs

Copy policy_template/, rename the policy directory and results JSON, and place your policy directly at the repository root:

<policy_id>/
├── README.md
├── leaderboard_<policy_id>.json
└── per_task_eval/
    └── <task_id>_<task_name>/
        └── <channel>/
            ├── per_episode.jsonl
            └── summary.json

The example logs mirror native Bench2Dex output, with null placeholders. Replace them with your complete sanitized evaluation files and remove unused example directories; do not fill the examples by hand.

Copy the README/results JSON from your GitHub submission. Use the same policy ID in the directory, filename, and model.id.

• Submit the original per_episode.jsonl and summary.json generated by Bench2Dex, after sanitization. Preserve the full native JSON structure, fields, and units. Do not convert logs to another schema or add template fields.
• Include all successful, failed, and error episodes from one selected run per task/channel. Report actual episode counts, budgets, and protocol changes.
• Submit at least Full results and logs; all four channels and SR, LSCR, Eff. are encouraged. List all 26 tasks in the results JSON, using null (NR) for missing scores.
• Results JSON channel keys are None, Equi., Inv., Full; directory names are None, Equi, Inv, Full. Retain the original channel metadata inside the native files.
• Only the leaderboard results JSON uses SR/LSCR percentages on 0–100 and Eff. as a ratio. Keep native log values unchanged; compute the published scores using the metric definitions.
• In the results JSON, use source.runId to identify the selected run and paths relative to <policy_id>/ for source.perEpisodeFile and source.summaryFile. Set evaluation.logs.repository to Bench2Dex/Leaderboard and policyPath to your policy ID. These submission references do not need to be added to the native logs.

🤝 Review

Open PRs against main here and in the GitHub leaderboard repository, then cross-link them. Policy IDs, scores, coverage, and protocol must agree. Authors may leave evaluation.logs.revision as null; after log review, maintainers record the accepted HF commit in the GitHub results JSON before publishing the entry.

🔒 Privacy and permissions

Replace private absolute paths with relative paths or consistent anonymized identifiers, including nested asset paths and error messages. Redact credentials, personal/confidential information, internal URLs, and restricted content in files, media, links, and PR descriptions. Preserve native fields, metrics, episode counts, seeds, task/channel identity, and randomization/protocol metadata. Confirm publication rights under applicable licenses and data-use terms.

📝 Citation

@article{yang2026bench2dex,
  title={Bench2Dex: Benchmarking Visuo-Tactile Bimanual Dexterous Manipulation Across Dexterous Hands},
  author={Yang, Zhenjie and Zhang, Yideng and Zhang, Dongjie and Jiang, Chenyu and Liu, Xianshuai and Li, Yufeng and Ge, Zuhao and Jiao, Xingyu and Zhang, Zheng and He, Kaiyu and others},
  journal={arXiv preprint arXiv:2609.15726},
  year={2026}
}