README.md

---

tags:

- robotics
- vision-language-action
- mobile-manipulation
- behavior-1k
- serf-vla
- pi0.5

---

# SERF-VLA BEHAVIOR-1K Checkpoints

Official checkpoint release for:

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

[[arXiv](https://arxiv.org/abs/2606.12956)] [[Website](https://existentialrobotics.org/serf/)] [[Code](https://github.com/ExistentialRobotics/SERF-VLA)]

## Overview

This repository contains PI0.5 baseline and SERF-VLA policy checkpoints for BEHAVIOR-1K experiments.

It includes:

* PI0.5 baseline checkpoints fine-tuned on individual BEHAVIOR-1K tasks
* SERF-VLA checkpoints conditioned on 4D spatiotemporal environment and robot feature maps

These checkpoints are for the policy learning component only. The SERF mapping component is not included in this repository.

## Checkpoint Initialization

All released checkpoints were initialized from the PI0.5 checkpoint pretrained on 50 BEHAVIOR-1K tasks:

* [`IliaLarchenko/behavior_50t_checkpoint`](https://huggingface.co/IliaLarchenko/behavior_50t_checkpoint)

This checkpoint was released by the first-place solution of the 2025 BEHAVIOR Challenge and was used as the initialization for both the PI0.5 baseline checkpoints and the SERF-VLA checkpoints in this repository.

## Checkpoints

| Folder                                                                  | Model          | Representation                       | Task        |
| ----------------------------------------------------------------------- | -------------- | ------------------------------------ | ----------- |
| `pi_behavior_b1k_fast--50t_lora--task-0021`                             | PI0.5 baseline | 2D image observation                 | 21 |
| `pi_behavior_b1k_fast--50t_lora--task-0022`                             | PI0.5 baseline | 2D image observation                 | 22 |
| `pi_behavior_b1k_fast--50t_lora--task-0026`                             | PI0.5 baseline | 2D image observation                 | 26 |
| `pi_serf_behavior_b1k_fast--4d_env_robot_feat_map--50t_lora--task-0021` | SERF-VLA       | 4D environment and robot feature map | 21 |
| `pi_serf_behavior_b1k_fast--4d_env_robot_feat_map--50t_lora--task-0022` | SERF-VLA       | 4D environment and robot feature map | 22 |
| `pi_serf_behavior_b1k_fast--4d_env_robot_feat_map--50t_lora--task-0026` | SERF-VLA       | 4D environment and robot feature map | 26 |

Each checkpoint follows the original policy checkpoint structure:

```text
checkpoint_name/
├── assets/
└── params/
```

## Usage

Download all checkpoints with `huggingface_hub`:

```python
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="byeonghyunpak/SERF-VLA",
    repo_type="model",
    local_dir="checkpoints/serf-vla-behavior-b1k",
)
```

To download a specific checkpoint folder only:

```python
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="byeonghyunpak/SERF-VLA",
    repo_type="model",
    local_dir="checkpoints/serf-vla-behavior-b1k",
    allow_patterns=[
        "pi_serf_behavior_b1k_fast--4d_env_robot_feat_map--50t_lora--task-0021/**"
    ],
)
```

For installation, data preparation, training, and evaluation instructions, please refer to the official code repository:

https://github.com/ExistentialRobotics/SERF-VLA

## 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 execution time.

## Citation

If you find these checkpoints useful, please cite our paper:

```bibtex
@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}
}
```

## Acknowledgements

This release builds on behavior-1k-solution, openpi, and BEHAVIOR-1K. We thank the authors and maintainers of these projects.