Action tokenizer that keeps movement rankings reports 94.8% on LIBERO

single source· 1 articles · confidence: low · first seen 2026-09-15 20:00 UTC

What this means for you

Nothing to migrate to: this is one arXiv preprint, with no weights, code or evaluation dates, and every number comes from a single Qwen3-VL-4B policy setup. If you train VLA policies, the transferable idea is the metric — reconstruction error alone can hide action adjustments that compression reversed.

A preprint presents an action tokenizer for vision-language-action models — systems that take camera images and instructions and emit robot commands as a sequence of discrete symbols. ActionPiece supervises representation learning and quantisation (mapping continuous movement onto a fixed set of code words) so that relative physical distances between movements survive compression. The paper argues that scoring commands separately with mean squared error misses adjustments that get flattened or reversed, and offers physical rank consistency to measure that. Under one Qwen3-VL-4B policy setup it reports 94.8% on LIBERO, 68.8% on unseen LIBERO-Plus, 71.9% on SimplerEnv and 51.5% across VLA-Arena L0–L2. No evaluation dates or baselines are given.

Key facts

  • ·ActionPiece reports 94.8% on LIBERO and 68.8% on unseen LIBERO-Plus, both under the same Qwen3-VL-4B policy training setup. source
  • ·The paper also reports 71.9% on SimplerEnv and 51.5% across VLA-Arena L0–L2. source
  • ·It introduces physical rank consistency, a metric for how well tokenisation preserves local physical distance rankings after reconstruction. source
  • ·ActionPiece jointly supervises representation learning and quantisation, with a frozen decoder executing the predicted tokens. source
  • ·No evaluation dates are given for the reported benchmark scores. source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire