HanLinqi/GR00T-N1.7-ForceBenchmark-4Tasks-H10-30K

GR00T N1.7 ForceBenchmark — four tasks, force / noforce

Eight final checkpoints trained for 30,000 steps each. Tasks: pickcube, stamp, whiteboard, weightsorting. Each task has force and noforce variants.

Files

Each complete checkpoint is in <task>/<variant>/checkpoint-30000/, including model shards, model and processor configurations, statistics, and saved training metadata.

Training

• Base model: GR00T-N1.7-3B
• Action horizon: 10
• Global batch size: 32
• Seed: 42
• State dropout probability: 0.0
• Padding enabled
• Training batch: formal_four_tasks_h10_20260913_234937
• Source code commit: 96fe97a8e983d8b8e98139d0f548876e5e745cca
• Completed: 2026-09-14 (UTC+08:00)
• Evaluation was not run for this batch; no evaluation success rates are claimed.

Download one checkpoint

from huggingface_hub import snapshot_download
snapshot_download(
    repo_id="HanLinqi/GR00T-N1.7-ForceBenchmark-4Tasks-H10-30K",
    allow_patterns="pickcube/force/checkpoint-30000/*",
    local_dir="checkpoints",
)

Use the corresponding ForceBenchmark GR00T implementation to load force-enabled models.