acceptodds
Under review as a conference paper at ICLR 2027

TopAct: Topology-Aware Action Tokens for Autoregressive Robot Policies

Abstract

An action tokenizer determines both the information retained about a robot trajectory and the categorical prediction problem presented to its policy. We introduce TopAct, an ordered codec that represents a 16-step action chunk with four finite-scalar-quantized tokens. Beyond reconstruction, its training objective aligns neighborhoods in action space with neighborhoods in latent space, before and after quantization. We evaluate matched topology-aware and reconstruction-only (Recon) codecs through action-only codec training, compact autoregressive policies, and an OpenVLA integration, holding the codec frozen while training each policy. TopAct preserves substantially stronger neighborhood structure after quantization and produces policy targets that are consistently easier to predict: token accuracy is higher for the compact policy on every suite, and under OpenVLA the TopAct arm converges faster and to a stronger token model. The same targets also support better closed-loop behavior: the compact policy reaches higher success on three of four suites, and the selected OpenVLA policy attains a higher success point estimate. Together, the findings motivate topology-aware action representations and the joint evaluation of codec quality, policy learning, and closed-loop behavior.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Action/latent representation, VLA/World Model, Topology-preservation, Inference acceleration

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_f3x3OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Paper violates the ICLR 2027 page limit policy and unfortunately should be rejected.

Critical strengths

Paper violates the page limit policy and unfortunately should be rejected.

Critical weaknesses

Paper violates the page limit policy and unfortunately should be rejected.

Questions for authors

Paper violates the page limit policy and unfortunately should be rejected.

Reviewer_VVZBOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposed an ordered action codec that compress a 16 step action chunk into four quantized tokens. The main contribution is a neighborhood-matching objective that aligns action smooth and similariy before and after quantization. The results show that the proposed method keeps the relation between actions but with higher reconstruction errors and limited success rate.

Critical strengths

I think the motivation to compress the action space is important and should be studied. The paper proposes TopAct, which compress action into four quantized tokens. The paper studies the relationship between reconstruction, generation, and the influence to downstream tasks, which I think is valuable. Reporting decoded action error alongside token accuracy is also helpful.

Critical weaknesses

The evidence is insufficiently robust to establish a reliable policy benefit. All codec and policy experiments use a single training seed. Under OpenVLA, TopAct achieves 40.8% versus 38.0% success, which the improvment is more or less incremental. A success-versus-training-step curve on independent evaluation states would support this claim more directly. The paper also lacks matched comparisons with OAT, FAST or VQVLA, please add more comparison and discussion with these action compression papers.

Questions for authors

see the weakness part and provide more evidence to support the effectiveness of the paper

Reviewer_9ByBOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper exceeds the maximum allowed page limit (the main text, including the Conclusion in Section 6, continues onto page 10). I therefore recommend desk rejection and have not assessed the technical content. I raised this with the AC in an earlier comment and am happy to revise my review if the AC decides the paper should be evaluated on its merits.

Critical strengths
Critical weaknesses
Questions for authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.