siyich/robovolo-robolab120

VoLoAgent Full — RoboLab120 evaluation outputs

Public RoboLab-format evaluation artifacts for VoLoAgent Full on all 120 RoboLab benchmark tasks, with 10 episodes per task for each language variant.

Results

Variant	Success	Episodes	Success rate
vague	354	1,200	29.50%
default	339	1,200	28.25%
specific	377	1,200	31.42%

Configuration

• VLA: pi0.5 DROID joint-position policy
• Mode: subgoal
• Failure monitor: VLM
• Recovery: replan tools
• VLM: aws/anthropic/bedrock-claude-opus-4-6
• Front camera enabled
• Grasp segmentation: SAM3
• Place segmentation: Molmo point
• Grasp top-down threshold: 0.85
• Check interval: 80 action chunks
• Subgoal timeout: 9999
• Frozen implementation commit: 59280ae5e2aff509111de5030c5e466fd7dc8307

Layout

Each variant is a standard RoboLab output folder:

<variant>/
├── episode_results.json
├── _batch_manifests/                 # original per-shard result manifests
└── <TaskName>/
    ├── data.hdf5
    ├── env_cfg.json
    ├── log_0.json ... log_9.json
    └── <instruction>_<episode>.mp4

Each task contains ten raw observation videos, ten viewport videos, and ten annotated videos.

Artifact audit

All variants contain 1,200 result rows, 120 task directories, ten per-episode JSON logs per task, and 30 videos per task. One source recording was absent after the completed run: default/ClearOrganicObjectsTask/data.hdf5. Its ten result rows, JSON logs, environment configuration, and all 30 videos are present. All other task HDF5 files are present.