Skip to content

Data Collection

Collect demonstrations for BEHAVIOR tasks using the JoyLo teleoperation system.

Preparation

This preparation step needs only be done once.

  • Make sure you are on the latest main branch of BEHAVIOR-1K
  • Update the latest robot assets 
    python OmniGibson/omnigibson/utils/asset_utils.py --update_omnigibson_robot_assets
  • Go inside 2025-challenge-task-instances and git pull
  • Clone 2026-challenge-task-instances into BEHAVIOR-1K/datasets

Outcome: You should see the following folders under BEHAVIOR-1K/datasets:

  • 2025-challenge-task-instances
  • 2026-challenge-task-instances
  • behavior-1k-assets
  • omnigibson-robot-assets

Data Collection Workflow

Step 1: Pull the Latest Code and Sampled Tasks

Pull the latest changes from both the main branch of BEHAVIOR-1K and datasets/2026-challenge-task-instances.

git pull

Step 2: Pick a Task

Available tasks can be found in 2026-challenge-task-instances/metadata/available_tasks.yaml.

Launch the following scripts to start data collection:

python joylo/scripts/launch_og.py --task-name TASK_NAME --recording-path HDF_PATH

python joylo/scripts/run_joylo.py

Step 3: Replay Trajectory

Run the following script to replay the trajectory. This will create one video.mp4 for visual QA and a qa_results.json, which includes the results from the QA script.

python joylo/scripts/replay_data.py HDF_PATH --task TASK_NAME --qa