AI/ robotics · ai · vla-models · tokenization

ActionPiece Tokenizer Hits 94.8% on Robot Benchmark LIBERO

ActionPiece, a new action tokenizer for robot AI, scored 94.8% on the LIBERO benchmark by preserving physical relationships between actions during compression.

A new way of turning robot movements into AI-readable tokens just posted strong scores on multiple robotics benchmarks.

Researchers built ActionPiece, a tokenizer that converts a robot's physical actions into discrete tokens for training vision-language-action models. Most tokenizers are graded on how closely a reconstructed action matches the original, a measure called mean squared error. The team argues that metric misses something: similar actions can get squashed into the same token, erasing the small adjustments that separate a good grasp from a dropped one. ActionPiece adds two training objectives that specifically preserve the relative distances between different actions, both when they're encoded and when they're assigned to a token.

Trained with the same Qwen3-VL-4B policy setup, ActionPiece hit 94.8% success on the LIBERO benchmark, 68.8% on the harder unseen LIBERO-Plus suite, 71.9% on SimplerEnv, and 51.5% across VLA-Arena's L0-L2 tasks. Those numbers matter because action tokenization is a quiet bottleneck in robot learning - get it wrong and a policy can look fine in testing while failing on tasks it has never seen.

The paper doesn't publish the baseline scores its comparison methods hit under the same setup, so there's no way to size up exactly how much of an edge this buys - a gap worth watching before anyone calls it a breakthrough.

TR

The Revision

Written by an AI system from the public sources credited above. How we write →