JIAlonglong/gr1-six-task-mini

GR1 Six Task Mini — v1 (pilot)

Six GR1 tabletop tasks, 20 fixed scenes each, 120 episodes. This is a small diagnostic benchmark, not the official UniT 24-task benchmark. No evaluation results are available yet.

Download

Download evaluation ZIP (no login required) · SHA256 checksum

Extract the archive to obtain gr1-six-task-mini/. It contains all 120 initial scenes, protocol, splits and evaluation tools, but not training demonstrations, simulator assets or model weights. Install simulator dependencies separately as described in the submission guide.

Contents

• protocol.json: frozen evaluation rules and scene digest.
• scenes.json: fixed scene identities, seeds and initial-state checksums.
• training-splits.json: 80 train / 20 val / 20 test demonstrations per task. Evaluation scenes are disjoint from these 720 episode IDs.
• initial-scenes/: first simulator state, scene metadata and XML extracted from the source demonstrations. No demonstration actions or future states.
• leaderboard.json, results/, videos/: verified-format submissions and fixed-scene videos when available.
• tools/: submission importer and portable scene loader. A model adapter and GR1 simulator/assets are required to execute evaluation.

Ranking

Only submissions with all 120 matching scenes completed are ranked, separately by training-data track. Main score is mean of six task success rates. Success means any simulator success within 720 control steps. Errors and timeouts cannot be omitted to improve the score. Scores are self-reported until independently reproduced. Single seed and 20 episodes/task give limited statistical precision.

Reuse and provenance

Derived from NVIDIA PhysicalAI-Robotics-GR00T-Teleop-Sim, © NVIDIA GEAR, CC BY-NC 4.0. This subset extracts initial states and defines new splits/protocol; it is not an official NVIDIA benchmark. Source simulator and assets retain their original licenses and must be installed separately. UniT reference: https://github.com/xpeng-robotics/UniT .

XML asset paths are rewritten by the scene loader for a user-supplied asset root; matching simulator/assets are required. Portability and complete closed-loop execution remain to be validated. Public protocol changes require a new version; do not silently replace v1 scenes.

Design and compute

See benchmark design and compute budgets. Optional tools/analyze_results.py reports per-task Wilson intervals and paired, task-stratified scene bootstrap. This is not across-training-seed uncertainty. Proposed robustness/latent diagnostics are clearly separated from implemented v1 metrics.

Latent action diagnostics

Optional diagnostics now include reconstruction RMSE/P95/tail error, per-layer discrete code usage/entropy, and predicted-vs-reference latent decoding gaps. diagnostics/ contains480 fixed validation windows and raw targets; these are explicitly separate from the120 closed-loop scenes. See LATENT_DIAGNOSTICS.md. No real model diagnostic results are available yet.

当前数据范围：仅人类遥操作采集的GR1机器人示范；不包含EgoDex等人类第一视角手部操作视频。当前成绩不能用于宣称人类视频到机器人的迁移收益。

Human-to-robot transfer track

The optional robot-human-transfer submission track pairs a robot-only baseline with480 additional EgoDex clips. Video-only and video+hand-pose settings are distinguished; equal GPU budget and equal robot exposure are separate comparisons. See HUMAN_TRANSFER.md. Only subset IDs/checksums are distributed, not EgoDex videos or converted annotations. No completed transfer results or validated mixed-training launcher are claimed.