YelinHe/behavior_2026_official_private_test

BEHAVIOR 2026 Official Private-Test Videos

Official results are available on the 2026 BEHAVIOR Challenge leaderboard.

Evaluation and recording

During the official private test, the organizers ran the simulator and accessed our inference service through its public IP endpoint. Their evaluation client sent observations to our service and executed the returned actions. This collection contains 2,192 videos across 100 tasks, recorded on our side from the observations received during those calls.

The client used K=10: each request supplied an observation and received 10 consecutive actions, which the client executed in order before sending the next observation. At the simulator's 30 Hz control rate, our service therefore received and recorded an observation approximately every 1/3 second, producing 3 fps videos. These are service-side observation recordings, not full 30 fps simulator videos. K refers to the number of actions returned per request, not the number of model predictions; a request can be served from an existing action buffer.

Video counts and interrupted runs

The official protocol evaluates 20 instances per task. Some official calls were interrupted and rerun, leaving recordings from both interrupted attempts and subsequent runs. Consequently, a task can have more than 20 videos, and the number of files does not equal the number of distinct evaluated instances. The collection retains the recordings available on our service, including incomplete attempts.

Contents

• videos/: observation recordings named {task}_t{task_id}_port{port}_r{rollout}_{id}.mp4.
• video_index.json: recording metadata, including task, port, rollout, timestamps, frame count, duration, and episode_closed_observed. This flag indicates whether an episode closure was observed by the recorder, not whether the task succeeded.
• robot/r1pro.yaml and wrapper/default_wrapper.py: the robot configuration and evaluation wrapper.

The collection does not include official instance IDs or per-episode Q-scores and success labels. Please refer to the official leaderboard for the reported evaluation results.