Figure 1: Trade-off between long-term foresight and short-term precision induced by single horizon	Figure 2: Overview of the proposed mixture-of-horizons strategy

• VLA models' performance is sensitive to the action chunk length (horizon). The single horizon induces an inherent trade-off between long-term foresight and short-term precision.

• We propose Mixture of Horizons (MoH), a plug-and-play strategy that fuses multiple horizons within a single policy to inherit the strengths of both with minimal training or inference overhead.

• MoH enables Dynamic Inference, selecting stable actions through cross-horizon consensus for higher efficiency and robustness.

Clone the repository and set up the conda environment:

git clone git@github.com:Timsty1/MixtureOfHorizons.git
conda create -n moh -y python=3.10
conda activate moh
pip install uv
cd MixtureOfHorizons
uv pip install -r requirements.txt
pip install packages/libero
pip install packages/openpi-client

This implementation requires modifying the transformers library to support PyTorch-type $\pi$ series models, which rely on gemma, paligemma, and siglip.

First, locate your conda environment path:

conda info --base

Then, copy the provided files to the transformers library directory (replace YOUR_CONDA_DIR with the path found above):

cp -r ./src/openpi/models_pytorch/transformers_replace/* YOUR_CONDA_DIR/envs/moh/lib/python3.10/site-packages/transformers/

1. Dataset: Download the LIBERO training set in LeRobot format from HuggingFace.
2. Base Models: Download the PyTorch base models for the $\pi$ series (converted from JAX) from Here.
3. Normalization Stats: Run the script 'scripts/compute_norm_stats.py' to compute normalization statistics, or use the pre-computed file provided in our Model Repo.

Before training, update the project path in your training scripts (train.py, train_pytorch.py, or train_pytorch_moh.py):

import sys
sys.path.append("YOUR_PROJ_DIR/MixtureOfHorizons/src")

Next, modify specific configurations (e.g., repo_id, pytorch_weight_path) in:

• scripts/run_train_moh.sh
• src/openpi/training/config.py

We trained our models on 4x A100 (80G) GPUs for 30k iterations with a batch size of 32. To start training with the MoH strategy:

bash scripts/run_train_moh.sh

We provide scripts for both standard evaluation and evaluation using Dynamic Inference. Our checkpoints are available at HuggingFace.

Prerequisites:

• Modify scripts/eval_on_libero.sh (or _dynamic.sh) to point to your YOUR_PROJ_DIR and correct checkpoint paths.

If you are running the evaluation for the first time in this conda environment, you must initialize the Libero simulator path configuration.

1. Run the test command manually in your terminal without piping to a log file.
2. When prompted: "Do you want to specify a custom path for the dataset folder? (Y/N):"
3. Type N and press Enter.
4. You can then safely interrupt the process (Ctrl+C).

Once configured, execute the evaluation scripts:

# Standard Evaluation
bash scripts/eval_on_libero.sh

# Evaluation with Dynamic Inference
bash scripts/eval_on_libero_dynamic.sh

Resources: Video records and raw logs of our experiments are available on Google Drive.

We express our gratitude to OpenPi, LIBERO, and RoboTwin for their open-source contributions.

If you feel that this paper, models, or codes are helpful, please cite our paper, thanks for your support!

@article{jing2025mixture_of_horizons,
  title={Mixture of Horizons in Action Chunking},
  author={Jing, Dong and Wang, Gang and Liu, Jiaqi and Tang, Weiliang and Sun, Zelong and Yao, Yunchao and Wei, Zhenyu and Liu, Yunhui and Lu, Zhiwu and Ding, Mingyu},
  journal={arXiv preprint arXiv:2511.19433},
  year={2025}
}