OAT: Ordered Action Tokenization

Authors: Chaoqi Liu, Xiaoshen Han, Jiawei Gao, Yue Zhao, Haonan Chen, Yilun Du · RSS 2026

Licences

Open Access
Paper: CC BY 4.0

Everything below draws on the full paper under its CC BY 4.0 licence; its figure is credited to the authors.

TL;DR

OAT compresses an action chunk into 8 ordered tokens, and any prefix decodes to a valid chunk. Autoregressive policies using it beat binning, FAST, QueST and diffusion policy.

Why it matters

Because any token prefix decodes to an executable chunk, one trained policy offers a trade-off between speed and precision at run time: on LIBERO, decoding 1 token takes 10.5 ms per inference and 8 tokens 27.4 ms on an NVIDIA A100, against 517.2 ms for binning. The ablations also show that token ordering, not compression alone, drives policy success, which is a concrete design rule for anyone building token-based robot policies.

Problem

Autoregressive robot policies need continuous actions turned into discrete tokens. Per-dimension binning produces hundreds of tokens per chunk, FAST (DCT plus byte-pair encoding) can generate token sequences that do not decode to a valid action chunk, and learned latent tokenizers such as QueST give compact but unordered tokens that suit next-token prediction poorly. The authors argue that no existing scheme offers all three of high compression, total decodability and left-to-right causal ordering.

Proposed solution

OAT is an autoencoder: a transformer encoder reads the action chunk together with learnable register tokens, finite scalar quantization (FSQ) discretizes the registers, and a transformer decoder reconstructs the chunk. Causal attention among the registers and nested dropout (keeping only a random number of leading tokens during training) make early tokens carry the coarse motion and later tokens add detail. Any prefix can then be decoded after padding the remaining positions with a learned mask token.

Key insights

Reconstruction error does not predict policy success: Bin and FAST reconstruct actions well but give the weakest autoregressive policies.
Ordering is the critical ingredient: the same tokenizer trained without nested dropout (OAT×) falls from 56.3% to 35.2% success on LIBERO, in some cases below QueST.
The decoder must accept every token sequence: FAST's byte-pair encoding can emit a coefficient count that does not match the action chunk, which cannot be reshaped into actions.
Bigger codebooks are not better: performance saturates at a moderate size and drops at the largest FSQ codebook, which the authors attribute to harder next-token prediction.
Early tokens capture the global motion and later tokens fine corrections; on the real robot, OAT with fewer than 4 tokens often got close to the goal but missed fine insertions.

Main results

LIBERO (libero10): OAT8 reaches 56.3% success vs QueST 48.2%, diffusion policy 36.6%, FAST 23.0% and Bin 14.4% (5 seeds × 50 rollouts per task).
RoboMimic: OAT8 73.1% vs diffusion policy 67.1% and QueST 66.9%. MetaWorld: OAT8 24.4% vs 19.3% and 17.9%. RoboCasa: OAT8 54.6% vs 54.0% and 52.3%.
Success rises with the number of decoded tokens, e.g. on LIBERO 11.7% (1 token), 39.8% (2), 46.4% (4) and 56.3% (8).
Latency per inference on one NVIDIA A100 (LIBERO): 10.5 ms with 1 token and 27.4 ms with 8, vs 517.2 ms for Bin (224 tokens) and 114.4 ms for FAST (44.2 tokens on average).
Real ARX-5 arm, 20 trials per task: OAT8 succeeded 16/20 on Pick & Place Ball and 16/20 on Stack Cups, vs diffusion policy 14/20 and 11/20, and QueST 11/20 and 8/20.
Removing nested dropout (OAT×) lowers success on all four benchmarks, e.g. RoboMimic from 73.1% to 61.1%.

Limitations

The number of decoded tokens is fixed at deployment; the authors note it should depend on each chunk's complexity and leave estimating this online as an open problem.
Short prefixes trade precision for speed: with fewer than 4 tokens the real robot often failed fine-grained insertions such as placing the ball fully into the cup.
Codebook size needs tuning: on LIBERO, success fell from 56.3% with FSQ levels [8,5,5,5] to 46.9% with [7,5,5,5,5] despite better reconstruction.
Real-world validation covers two tabletop tasks on one fixed-base arm with a single webcam, using 200 demonstrations and 20 trials per task.
Absolute success stays low on some benchmarks (best MetaWorld result 24.4%), and on RoboCasa OAT8 (54.6 ± 1.1%) is close to diffusion policy (54.0 ± 1.6%).

Field impact

Finalist for the RSS 2026 Outstanding Paper Award, with 27 citations. The paper links a project website with interactive reconstructions (ordered-action-tokenization.github.io); it does not state a code release.

Abstract

Autoregressive policies offer a compelling foundation for scalable robot learning by enabling discrete abstraction, token-level reasoning, and flexible inference. However, applying autoregressive modeling to continuous robot actions requires an effective action tokenization scheme. Existing approaches either rely on analytical discretization methods that produce prohibitively long token sequences, or learned latent tokenizers that lack structure, limiting their compatibility with next-token prediction. In this work, we identify three desiderata for action tokenization - high compression, total decodability, and a left-to-right causally ordered token space - and introduce Ordered Action Tokenization (OAT), a learned action tokenizer that satisfies all three. OAT discretizes action chunks into an ordered sequence of tokens using transformer with registers, finite scalar quantization, and ordering-inducing training mechanisms. The resulting token space aligns naturally with autoregressive generation and enables prefix-based detokenization, yielding an anytime trade-off between inference cost and action fidelity. Across more than 20 tasks spanning four simulation benchmarks and real-world settings, autoregressive policies equipped with OAT consistently outperform prior tokenization schemes and diffusion-based baselines, while offering significantly greater flexibility at inference time.

Open Source Robotics