BEHAVIOR-1K v3.7.1

In this release, we fixed various bugs and added more features/supports for the 1st BEHAVIOR challenge. Visit https://behavior.stanford.edu/ for more information!

Updates

• Removed grasp_left and grasp_right from robot proprioception and updated the demo dataset accordingly.
  • Now, the dimension of observation.state in the dataset is 256 instead of 258
• Updated task-instances dataset, included 2 more metadata files for data replay
• Fixed library errors (av, gspread, pandas, cffi) during installation
• Added RGBLowResWrapper and set it to be the default wrapper during evaluation rollout
• Added GOP-level randomization to BehaviorLeRobotDataset for faster data access during training
• Added point cloud modality to vision sensor
• Updated camera relative pose to base implementation, changed n_render_iterations to 1 during evaluation rollout
• Fixed bugs in demo replay script
• Added NO_OMNI_LOG macro
• Updated challenge rules and documentations:
  • Changed evaluation rollout number from 20 * 5 to 10 * 1
  • Specified robot base global pose is not allowed in standard track
  • Added video recordings as part of the submission

To migrate from v3.7.0 to v3.7.1

• If you already downloaded the demo dataset, please update this by syncing local dataset with HF

  ds = BehaviorLeRobotDataset(
      repo_id="behavior-1k/2025-challenge-demos",
      root=PATH_TO_DATASET_ROOT,
      tasks=LIST_OF_TASK_NAMES,
      download_videos=False,
      local_only=False,
      force_cache_sync=True
  )
  

  Specify task names under tasks, or set to None if you want to update all tasks.
  If you are only downloading the dataset, you can safely exit the program after seeing something like

  Generating train split: 23346708 examples [03:25, 112135.76 examples/s]
  

• Pull the latest code from v3.7.1 (79f0f5), then within the behavior conda environment run:

  ./setup.sh --bddl --omnigibson --dataset
  

Full Changelog: v3.7.0...v3.7.1

What's Changed

• Fix: broken links on b100 doc by @hang-yin in #1820
• Update eval by @wensi-ai in #1821

Full Changelog: v3.7.0...v3.7.1