ActionPiece: Rethinking Action Tokenization for Autoregressive Vision-Language-Action Models
ActionPiece improves action tokenization for vision-language-action models via physical rank consistency, reaching 94.8% on LIBERO with Qwen3-VL-4B.
The paper introduces physical rank consistency (PRC), a metric measuring whether tokenization preserves local physical distance rankings of actions after reconstruction. ActionPiece preserves physical action relationships through joint supervision of representation learning and quantization, alongside reconstruction losses. Under the same Qwen3-VL-4B policy training setup, ActionPiece achieves 94.8% on LIBERO, 68.8% on unseen LIBERO-Plus, 71.9% on SimplerEnv, and 51.5% across VLA-Arena L0-L2.
- Pointwise MSE fails to capture distortion of relative action adjustments.
- PRC supervises near-far ordering in encoder and quantized feature distances.
- Quantization regularization applies the same ordering to codeword assignments.
- Ablations show both objectives jointly improve PRC and policy success.
- Products
- ActionPiece
- AI models
- Qwen3-VL-4B
From the original article
1 min read at huggingface.co
Action tokenizers play a central role in autoregressive vision-language-action (VLA) models, determining both the targets for policy training and the executable commands recovered from predicted tokens.
By Shijie Lian, Bin Yu, Zhaolong Shen · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.