SERF-VLA

Official source code repository for:

SERF: Spatiotemporal Environment and Robot Feature Map for Long-Horizon Mobile Manipulation

[arXiv] [Website] [Video]

This repository contains the policy learning code for SERF-VLA on BEHAVIOR-1K, covering:

• Preparing BEHAVIOR-1K data
• Training VLA policies
• Evaluating policies on BEHAVIOR-1K

Note: this repository does not include the mapping component of SERF.

Repository Status

Released:

• Setup instructions
• Policy training and evaluation code

Coming soon:

• Fine-tuned PI0.5 checkpoints
• SERF-VLA checkpoints

Installation

See docs/INSTALLATION.md for environment setup and BEHAVIOR-1K installation instructions.

Data

This repository assumes access to BEHAVIOR-1K data and task assets. Dataset preparation instructions are provided in docs/DATASET_PREPARATION.md. SERF map assets should be generated with SERF-mapping and placed under datasets/SERF-BEHAVIOR-1K-MAP.

Expected data layout:

datasets/
  2025-BEHAVIOR-1K-CHALLENGE/
    data/
      task-0021/
      task-0026/
      ...
  SERF-BEHAVIOR-1K-MAP/
    exported_neural_points/
    map_models/

Checkpoints

Before training SERF-VLA policies, download the PI0.5 checkpoint pretrained on the 50 tasks from the 2025 BEHAVIOR Challenge. We use the checkpoint released by the first-place challenge solution, behavior-1k-solution, as the initialization for our experiments.

Run the following command from the repository root:

uv run python - <<'PY'
from huggingface_hub import snapshot_download

ckpt_dir = "checkpoints/behavior-1k-solution"

snapshot_download(
    repo_id="IliaLarchenko/behavior_50t_checkpoint",
    repo_type="model",
    local_dir=ckpt_dir,
)

print(f"Checkpoint downloaded to: {ckpt_dir}")
PY

This places the checkpoint under checkpoints/behavior-1k-solution, matching the default paths used by the training and evaluation configs.

SERF-VLA checkpoints and our fine-tuned PI0.5 checkpoints will be added to this section when they are ready for release.

Training

Use the wrapper scripts in scripts/train from the project root. Each script selects the corresponding training preset and accepts common overrides such as --task-id, --batch-size, and --num-train-steps.

For the reported experiments, we fine-tune each model for 20k steps with a batch size of 16. Training was run on one NVIDIA H100 GPU.

Example usage:

bash scripts/train/train_2d_image.sh --task-id 0021
bash scripts/train/train_3d_env_feat_map.sh --task-id task-0021
bash scripts/train/train_4d_env_feat_map.sh --task-id task-0021
bash scripts/train/train_4d_env_robot_feat_map.sh --task-id task-0021

Run any script with --help to see its available arguments.

Evaluation

Use the wrapper scripts in scripts/test from the project root. These scripts start the policy server, wait for it to initialize, and then launch the BEHAVIOR-1K / OmniGibson evaluation script.

Note: BEHAVIOR-1K evaluation is non-deterministic. Results can differ across repeated runs due to variability in the underlying physics simulation and error accumulation over long-horizon rollouts.

Runtime: Evaluation is computationally expensive. A single episode can take several hours, and a full 20-episode task evaluation may take several days to complete.

Speed settings: If evaluation is too slow, we recommend disabling rollout video and per-step Q-score logging with --write-video false, --write-third-person-video false, and --record-step-q-score false. For task 26, --record-step-q-score false is required because per-step Q-score computation is too slow.

Setup Evaluation Script

Copy the files in src/serf_b1k/learning to BEHAVIOR-1K/OmniGibson/omnigibson/learning.

# Execute from the project root
cp -r src/serf_b1k/learning/* BEHAVIOR-1K/OmniGibson/omnigibson/learning/

Patch Task 21 Goal

For task 21 (collecting_childrens_toys), SERF evaluation expects all dice, teddy bears, board games, and train sets to be inside the bookcase. The evaluation wrappers patch the corresponding BDDL goal automatically before running task 21. To apply or verify the patch manually, run:

python scripts/setup/patch_collecting_childrens_toys_bddl.py
python scripts/setup/patch_collecting_childrens_toys_bddl.py --check

Example usage:

bash scripts/test/test_2d_image.sh \
  --task-id 0021 \
  --checkpoint-path exps/path/to/checkpoint

bash scripts/test/test_3d_env_feat_map.sh \
  --task-id 0021 \
  --checkpoint-path exps/path/to/checkpoint

bash scripts/test/test_4d_env_feat_map.sh \
  --task-id 0021 \
  --checkpoint-path exps/path/to/checkpoint

bash scripts/test/test_4d_env_robot_feat_map.sh \
  --task-id 0021 \
  --checkpoint-path exps/path/to/checkpoint

For the pretrained 2D baseline, use:

bash scripts/test/test_2d_image_pre.sh --task-id 0021

Run any evaluation wrapper with --help for options such as video logging, map dataset paths, robot map paths, and pass-through OmniGibson overrides.

Citation

If you find this repository useful, please cite our paper:

@article{kim2026serf,
  title = {SERF: Spatiotemporal Environment and Robot Feature Map for Long-Horizon Mobile Manipulation},
  author = {Kim, Sunghwan and Pak, Byeonghyun and Long, Kehan and Tian, Yulun and Atanasov, Nikolay},
  journal = {arXiv preprint arXiv:2606.12956},
  year = {2026}
}

License

This project is released under the license provided in LICENSE.

Acknowledgements

This repository is primarily based on behavior-1k-solution and openpi. We thank the authors and maintainers of these projects, as well as the BEHAVIOR-1K team.