BEAST

An action tokenizer that compresses a chunk of actions into a fixed number of tokens using B-spline control points.

BEAST is an action tokenizer proposed in June 2025 by a team at the Karlsruhe Institute of Technology (KIT) in Germany, published at NeurIPS 2025. The idea is to fit an action sequence with a B-spline curve (a smooth curve whose shape is set by a small number of “control points”), then treat the control points as tokens: they can be quantized into discrete tokens for a language model, or kept as continuous values. Compared with tokenizers like FAST, which need separate training and produce variable-length output, BEAST needs no training and always produces a fixed number of tokens per chunk, so decoding can happen in parallel in a single forward pass; the B-spline also guarantees smooth transitions between the start and end of adjacent action chunks, reducing jitter. The paper plugs it into small models like Florence-2 (BEAST-F) and into architectures like ACT, gets competitive results on CALVIN and LIBERO, and reaches roughly twice π0's inference throughput.

ExampleBEAST-ACT turns the 100-step action sequence ACT would normally predict step by step into just 15 control points, which a B-spline then reconstructs into a smooth trajectory.

Also called

B-spline Encoded Action Sequence Tokenizer, BEAST-F, BEAST-ACT

Related

Action Tokenizer · Action Chunking · π0-FAST · Parallel Decoding · Florence-2 · Action Chunking with Transformers

Sources

BEAST: Efficient Tokenization of B-Splines Encoded Action Sequences for Imitation Learning (arXiv 2506.06072)

As of

2025-10