Summary
Introduces “physical rank consistency” (PRC) as a criterion for action tokenizers — measuring whether tokenization preserves the relative physical-distance ordering of actions after reconstruction, not just pointwise MSE — and proposes ActionPiece, which jointly supervises representation learning and quantization to preserve this ordering.
Key Contributions
- New evaluation criterion (physical rank consistency) that goes beyond standard MSE-based reconstruction metrics for action tokenizers
- Joint training objective combining physical-rank preservation in the encoder with matching ordering constraints in the quantization/codeword space
- State-of-the-art aggregate results on SimplerEnv (71.9%) and VLA-Arena (51.5%), plus strong LIBERO/LIBERO-Plus numbers
Strengths
- The critique of MSE as an insufficient tokenizer-quality metric is well-motivated and could influence how the field evaluates future tokenizers
- Broad benchmark coverage (LIBERO, LIBERO-Plus, SimplerEnv, VLA-Arena) supports the generality of the claimed gains
Weaknesses
- “Physical rank consistency” as a proxy for real physical fidelity still doesn’t guarantee low-level control precision is preserved; it’s an ordinal, not metric, guarantee
- Unclear how sensitive gains are to the choice of quantization codebook size
Open Questions
- Does PRC-optimized tokenization transfer benefits to continuous/flow-based action heads, or is it specific to autoregressive discrete-token VLAs?
- How does the method perform on high-DoF dexterous/bimanual action spaces where physical distance is harder to define?
Significance
Action tokenization design has lagged behind backbone/architecture research in VLA papers; this contributes a more principled quality criterion that could become a standard tokenizer benchmark.