ActionPiece: Rethinking Action Tokenization for Autoregressive Vision-Language-Action Models
The paper argues that action tokenizers need to preserve physical relationships between motions, not just minimize reconstruction error.
The authors introduce physical rank consistency, a metric for checking whether reconstructed actions keep the same near-far ordering as the original actions. Their ActionPiece tokenizer adds supervision for those physical distance relationships during representation learning, quantization, and reconstruction. In the reported Qwen3-VL-4B setup, it reaches 94.8% on LIBERO, 68.8% on unseen LIBERO-Plus, 71.9% on SimplerEnv, and 51.5% across VLA-Arena L0-L2. Ablations in the paper attribute the gains to the two added objectives improving both PRC and policy success. HF Daily Papers' note
The authors introduce physical rank consistency, a metric for checking whether reconstructed actions keep the same near-far ordering as the original actions. Their ActionPiece tokenizer adds supervision for those physical distance relationships during representation learning, quantization, and reconstruction. In the reported Qwen3-VL-4B setup, it reaches 94.8% on LIBERO, 68.8% on unseen LIBERO-Plus, 71.9% on SimplerEnv, and 51.5% across VLA-Arena L0-L2. Ablations in the paper attribute the gains to the two added objectives improving both PRC and policy success. HF Daily Papers' note
score 5