BEAST: Efficient Tokenization of B-Splines Encoded Action Sequences for Imitation Learning

Hongyi Zhou1, Weiran Liao2, Xi Huang3, Yucheng Tang4, Fabian Otto5, Xiaogang Jia 1,2, Xinkai Jiang1,2, Simon Hilber 1, Ge Li1,2, Qian Wang2, Ömer Erdinç Yağmurlu1, Nils Blank1, Moritz Reuss1, Rudolf Lioutikov1
1. Intuitive Robots Lab (IRL), KIT, Germany
2. Autonomous Learning Robots (ALR), KIT, Germany
3. Intelligent Process Automation and Robotics Lab (IPR), KIT, Germany
4. Institute for Robotics and Autonomous Systems (IRAS), HKA, Germany
5. Microsoft Research, UK
Paper Hugging Face Code
NeurIPS 2025

Abstract

We present the B-spline Encoded Action Sequence Tokenizer (BEAST), a novel action tokenizer that encodes action sequences into compact discrete or continuous tokens using B-spline. In contrast to existing action tokenizers based on vector quantization or byte pair encoding, BEAST requires no separate tokenizer training and consistently produces tokens of uniform length, enabling fast action sequence generation via parallel decoding. Leveraging our B-spline formulation, BEAST inherently ensures generating smooth trajectories without discontinuities between adjacent segments. We extensively evaluate BEAST by integrating it with three distinct model architectures: a Variational Autoencoder (VAE) with continuous tokens, a decoder-only Transformer with discrete tokens, and Florence-2, a Vision-Language Model with an encoder-decoder architecture, demonstrating BEAST's compatibility and scalability with large pretrained models. We evaluate BEAST across three established benchmarks consisting of 166 simulated tasks and on three distinct robot settings with a total of 8 real-world tasks. Experimental results demonstrate that BEAST (i) significantly reduces both training and inference computational costs, and (ii) consistently generates smooth, high-frequency control signals suitable for continuous control tasks while (iii) reliably achieves competitive task success rates compared to state-of-the-art methods.

BEAST Tokenizer

Efficient Tokenization without Training

Architecture Overview

B-Splines as an Efficient Action Representation

Architecture Overview

BEAST-F: A tiny (0.77B) yet powerful VLA building upon BEAST Tokenizer

Architecture Overview

Getting Started

We provide a Hugging Face implementation for easy access. Simply install the dependencies and start using BEAST in just a few lines of code.

Installation

pip install torch numpy matplotlib einops transformers

Note: CUDA is recommended for optimal performance, but CPU is also supported by setting device="cpu".

Quick Start

from transformers import AutoProcessor
import torch

# Initialize the BEAST processor
beast = AutoProcessor.from_pretrained(
    "zhouhongyi/beast",
    trust_remote_code=True,
    num_dof = 3,
    num_basis = 20,
    seq_len = 50,
    degree_p = 3,
    device = 'cpu'
)

# Create random trajectory data: 10 trajectories, 50 time steps, 3 dimensions
trajectories = torch.randn(10, 50, 3)

# Encode trajectories into discrete tokens
tokens = beast.encode_discrete(trajectories, update_bounds=True)
print(f"Encoded tokens shape: {tokens.shape}")

# Decode tokens back to continuous trajectories
reconstructed_trajectories = beast.decode_discrete(tokens)
print(f"Reconstructed trajectories shape: {reconstructed_trajectories.shape}")

# Calculate reconstruction quality
mse_loss = torch.mean((trajectories - reconstructed_trajectories) ** 2)
print(f"MSE Loss: {mse_loss.item()}")

Calvin

LIBERO

Aloha_Sim

Insertion

Transfer

Franka Challenge

Fold

Mixer

Pour

Sweep

Franka Kitchen

Fold

Mixer

Pour

Real World Aloha

BibTeX

@inproceedings{
zhou2025beast,
title={{BEAST}: Efficient Tokenization of B-Splines Encoded Action Sequences for Imitation Learning},
author={Hongyi Zhou and Weiran Liao and Xi Huang and Yucheng Tang and Fabian Otto and Xiaogang Jia and Xinkai Jiang and Simon Hilber and Ge Li and Qian Wang and {\"O}mer Erdin{\c{c}} Ya{\u{g}}murlu and Nils Blank and Moritz Reuss and Rudolf Lioutikov},
booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems},
year={2025},
url={https://openreview.net/forum?id=rQCl1sf62w}
}