Rollouts & evidence
See the task interactions, then preserve the artifacts needed to explain a policy score.
Teleoperation previews.
These videos illustrate demonstration collection. They are not scored baseline-policy rollouts and do not substantiate a leaderboard number.
What to retain for each episode
A score should be traceable to a task, seed, scene, policy and terminal outcome. Save the stage transition trace for long-horizon episodes so a failed sequence can be diagnosed independently of its aggregate result.
{
"env_id": "SharpaAngledClickMouse-v1",
"seed": 10001,
"scene_background": "robocasa_l3_s2",
"checkpoint_hash": "<record actual hash>",
"adapter_revision": "<record actual revision>",
"success": false,
"control_steps": 500,
"stage_trace": [],
"video_path": "<episode video>",
"log_path": "<episode log>"
}Suggested documentation record only, not a universal implemented output schema. Add the registered horizon, task fingerprint, observation configuration and background manifest from your runner.
Scored policy rollouts
Scored baseline videos will be added when policy, checkpoint, seeds and run logs are supplied together. Collection previews remain explicitly separate from evaluation evidence.