ROBOT EVALUATION ARCHIVE
Watch the policy.
See every outcome.
A coding agent controlling robots through visual feedback. Browse recorded rollouts, from the first move to the final result.
Separate benchmarks, different tasks and control protocols. Success rates are reported independently.
THE COMPLETE RECORD
The rollout collection.
Loading recorded episodes…
Loading tasks…
No matching tasks.
Try another search or clear the filters.
Unable to load the collection. Please refresh the page. For local use, serve this folder with a static web server.
HOW TO READ THIS GALLERY
The task is visual.
The score is native.
These evaluations use GPT-6-Astra with high reasoning. Each episode starts with a fresh context and receives its native task instruction, RGB camera views and robot state. Camera calibration follows each benchmark’s protocol; RoboDojo exposes no calibration or depth.
Read the evaluation protocolAll scored episodes are included
The published selection includes successful and unsuccessful episodes. Source selection and infrastructure retries are documented in the evaluation protocol.
Success comes from the simulator
Results use each benchmark’s native success condition, not the agent’s written assessment. The policy receives no object ground-truth poses, reward or success labels, training demonstrations, or cross-episode memory. In interactive RoboDojo tasks, the native support arm’s demonstration remains visible as part of the task.
Playback and control time are different
The player describes each benchmark’s recorded cameras and playback rate. Thinking time is omitted from video playback.
Sampling and comparison
Sampling is reported separately for each published benchmark. Different tasks and protocols prevent direct comparison or pooling across benchmarks.