acceptodds
Under review as a conference paper at ICLR 2027

Same States, Different Actuators: What Single-Actuator Data Must Cover for Compositional World Models

Abstract

Compositional world models aim to predict unseen actuator combinations without collecting data for every combination. For torque-driven systems, the joint response at a fixed state is the sum of single-actuator responses up to a residual of 1.7–13.6% of the actuated response, so single-actuator data suffice if each actuator's response is known at the states the combination visits. Reset-only collection observes each actuator only at states reached by that actuator alone from reset, and existing coverage measures cannot flag this: concentrability is unbounded for joint actions under any single-actuator data, and state coverage ignores which actuator is observed at which state, the dataset's actuator–state pairing. A matched-pool intervention changes which actuator is executed at each state of a fixed pool, keeping per-actuator action marginals fixed, and so isolates the pairing. In a linear switched class, we prove that histories engaging only a proper subset of actuators, as reset-only histories do, can leave joint behaviour undetermined even though together they reach every state, and we give a necessary and sufficient condition for component observations to determine it. A collector must therefore let a history engage several actuators; state-carrying collection does so by switching actuators without resetting. Across four MuJoCo environments and two model classes, balanced pairing reduces one-step prediction error on unseen combinations by 8.3–29.2%, and by 8.4–25.4% when the joint prediction is assembled from the model's own single-actuator responses. At the same interaction budget, state-carrying collection reduces one-step error by 6.3–34.6% relative to reset-only collection, and its advantage extends to 120 and 98 of 121 unseen actuator subsets for the two model classes and to ten-step rollouts. The artifacts are available at https://figshare.com/s/2aec1f48b5d5a038aca6.

Primary area:
learning on time series and dynamical systems
Keywords:
compositional generalization, data collection, learned dynamics models, world models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_PWRBOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper builds on the key insight that for compositional world models, state coverage is not enough, actuators needs to be observed in relevant state contexts. Through scientific studies through the problem of single-actuator data, this paper shows that changing actuator–state pairings while holding the state pool and per-actuator action marginals fixed improves prediction of unseen actuator combinations. This is further accompanied with a theoretical argument on linear switch-system class and proposes a data-collection strategy that switches actuators without resetting the state, improving compositional prediction at a matched interaction budget.

Critical strengths
  1. It studies an interesting problem of how to collect data in a single-actuator regime such that the learned world model can generalize better to unseen state-action combinations.
Critical weaknesses
  1. Writing is hard to follow. The introduction does not provide clear enough motivation for the project, and the related works does not provide enough context for readers to understand the problem setting.
Questions for authors
  1. Could the authors provide a targeted matched-pool and reset-versus-carried using additive / action-affine learner for comparison under the final experimental protocal?
Reviewer_4Mi3OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how a world model can predict unseen multi-actuator joint actions when data collection is restricted to actions with at most one nonzero actuator input per timestep. The authors propose an actuator–state pairing perspective, arguing that datasets with the same pooled state distribution can support different levels of compositional prediction accuracy because their actuator–state pairings differ. CARRIED achieves lower prediction error on most evaluated unseen actuator subsets, with gains also extending to ten-step open-loop rollouts.

Critical strengths
  1. Within the authors' given setting, several interesting theoretical analyses are conducted.
  2. Multi-faceted experimental and diagnostic validation is employed.
Critical weaknesses
  1. The most confusing aspect of this paper, in my view, is its motivation. Why would the environment described in the paper arise in practice? The cited work [1] appears to require all other joints to be held at fixed positions, which is fundamentally different from the setting in this paper where at most one actuator input is nonzero at each timestep.
  2. The connection between theoretical results and the proposed data collection algorithm is weak. The theoretical analysis characterizes when exact component observations determine trajectories under every switching or simultaneous actuator schedule within a restricted linear switched class. However, the CARRIED collection protocol redraws the active actuator uniformly with replacement at fixed intervals while retaining the current physical state, without directly using the composition certificate to guide actuator selection or actively targeting undetermined response directions. Although Appendix G.3 reports an exploratory comparison with a certificate-inspired rule, the connection between the composition certificate and the main collection protocol remains qualitative.

[1] Zhiqiang Wang et al. Research on identification of minimum parameter set in robot dynamics and excitation strategy.

Questions for authors
  1. Could the authors further clarify the motivation of this paper? This is critical to the final review decision. What specific scenarios give rise to the single-actuator constraint? Please provide a concrete explanation of which practical data collection conditions require the torque inputs of all remaining actuators to be zero at each timestep.
  2. While the matched-pool experiments isolate the effect of actuator–state pairing, the RESET-to-CARRIED comparison changes both the pairings and the visited state distribution. Could the authors further disentangle these two factors and quantify their respective contributions to the observed improvement in compositional prediction?
Reviewer_U3gaOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

The paper studies single-actuator data collection for learned dynamics models that must predict joint actuation. It shows that balanced actuator-state pairing improves prediction (matched-pool experiment), gives identifiability results for a linear switched system, and proposes collection that switches actuators within an episode.

Critical strengths

The topic is important: data collection for models that must generalize to unseen action combinations. The theory appears correct, and the experiments appear sound, although the paper is hard to follow.

Critical weaknesses
  1. Venue fit. The core content is identification of switched and bilinear systems, which is standard at HSCC, L4DC, and CDC. Reviewers there can assess the contribution against the state of the art better than ICLR reviewers.

  2. Limited novelty. The paper does not cite the relevant literature: realization theory, identifiability, and persistence of excitation for switched and bilinear systems (Petreczky; Petreczky, Bako, and van Schuppen, HSCC 2010), and simultaneous versus separate excitation (Gevers et al., Automatica 2006). A quick LLM-guided literature search found these references. In that literature, Theorem 1 is an instance of a known fact: switching sequences must excite all relevant mode products. Theorem 2 adds PSD bounds, which is a small extension. The practical insight reduces to standard excitation advice: choose the actuator independently of the state (Appendix D.3).

  3. Superficial world-model framing. Both learners are full-state, one-step regressors. The paper excludes pixels, partial observation, and planning. The downstream control results are null or mostly not significant (Appendix G, Table 15). The paper also gives no world-model setting where the one-actuator constraint applies.

Questions for authors
  1. What concrete world-model setting requires single-actuator collection?
  2. Which results are new relative to switched and bilinear system identification?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.