acceptodds
Under review as a conference paper at ICLR 2027

PileBelief: Persistent Physical State for Interaction-Driven World Modeling

Abstract

World models allow robots to anticipate action consequences before execution. This capability is especially valuable in excavation, where each scoop reshapes the terrain and affects subsequent actions. Local observations, however, cannot fully reveal the support and material-response conditions governing excavation outcomes. We present PileBelief, an interaction-driven persistent world model for partially observed excavation that retains physical evidence beyond the visible surface. It combines an observation-conditioned physical prior with world-addressed deformation memory and physical-response memory. Action-aligned reads and gated residual corrections refine terrain-change and outcome predictions. Completed interactions update measured belief, while hypothetical actions advance a separate imagined state with fixed deployment weights. Experiments show that interaction history resolves current-view ambiguity, remains useful across delayed revisits, and improves observation-free multi-step prediction. Offline candidate ranking reduces action-selection regret by 65.5% relative to the current-observation-only Local baseline. Target-domain studies in Newton/MPM and real excavation demonstrate the value of interaction history across dynamics and sensing conditions. These results identify persistent physical belief as a useful representation for world models of environments that robots continually reshape.

Primary area:
applications to robotics, autonomy, planning
Keywords:
World Models, Persistent Memory, Physical State Representation, Robotic Excavation, Partial Observability

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_6GkrOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies how to predict the result of a digging action before it is executed, including how the pile will change and how much material the bucket will collect, so that a robot can choose better actions. The main difficulty is that the robot sees only part of the pile, and areas that look the same can respond differently because of earlier digging or hidden soil conditions. The paper proposes PileBelief, a model that records where the pile has changed and how it responded to past actions, and uses this record to improve its predictions. PileBelief outperforms a model that uses only the current view, both in simulation and on real excavation data. It has not yet been tested in real-time control of a physical machine.

Critical strengths

Excavation is a strong and well-chosen case for world models in environments that the robot itself reshapes, and the motivation is sound. The proposed method is interesting, separating memory of where the pile changed from how it responded.

Critical weaknesses
  1. Several key terms are not clearly defined: Newton/MPM is never explained, the belief state is not defined in one place and varies across experiments, and the resampling operator is only named.
  2. The problem formulation is unclear, since the model's inputs and outputs are not clearly specified and differ across domains.
  3. The paper does not show the model's value at the task level. It does not define a task-level objective.
  4. The paper lacks heuristic baselines, such as selecting the candidate with the largest swept volume.
Questions for authors
  1. What exactly are the model's inputs and outputs in each domain?
  2. How does PileBelief compare with a heuristic baseline, such as selecting the candidate with the largest swept volume?
Reviewer_BHNWOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors present PileBelief, an interaction-driven persistent world model for partially observed excavation that retains physical evidence beyond the visible surface. It stores terrain deformation in registered world coordinates and recent physical outcomes, then uses this evidence to refine predictions for candidate actions. The method is evaluated in experiments using a custom simulator (Netwon/MPM) and in a small real excavation dataset.

Critical strengths

The problem is well motivated: Local observations may omit physical information revealed by earlier scoops. The authors introduce an interaction-driven world model that aims to resolve this local-observation ambiguity.

Critical weaknesses

The paper addresses a very specific problem, and therefore the contribution appears to be highly application-specific. I could envision that similar concepts would be applicable to additional problems. However, the paper targets specifically a single particular application, i.e. excavation, without any attempt to present a more general formulation. Therefore, the paper does not seem to establish a broadly applicable ML method or a conceptual advance beyond existing excavation predictors and world models.

The paper does not include any theoretical results or in-depth analysis. It could be more appropriate for specific robotics venues.

Questions for authors

Q1: What is the methodological novelty over existing excavation world models and spatial-memory architectures, beyond combining deformation and response history?

Q2: Can you report uncertainty as part of the results?

Reviewer_6K9aOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PileBelief integrates "observation-based local predictors" with "recorded deformation history" and "physical response history." Through action-aligned read operations, the system generates bounded, gated residual corrections. Actual interactions update the observed state, while hypothetical rollouts update a virtual copy of that state. In controlled excavation experiments, the study compares several approaches: baselines using recurrent models and attention mechanisms, memory-based interventions, delayed revisits, free-form rollouts, and offline candidate ranking. Furthermore, the study evaluates implementations trained independently within the target domain, utilizing both simulations based on Newtonian mechanics—specifically the Material Point Method (MPM)—and real-world excavation experiments.

Critical strengths
  1. The issue of partial observability is both concrete and scientifically grounded; identical height maps can potentially mask distinct material responses.
  2. The supplementary evaluation section is notably candid in defining the scope of the research. The analysis concerning the material-holdout set reports interval-based results at the population level, which encompass improvements that remain subject to uncertainty regarding volumetric and force-response characteristics.
Critical weaknesses
  1. Concepts such as spatial mapping, recurrent memory, history-based adaptation, and gated residuals are already well-established; CoDeGa, for instance, has previously explored how to leverage completed exploration segments ("scoops") to adapt to environmental shifts (material shifts). Table 1 indicates that this method achieves a 3.6% improvement in five-step error compared to the History Transformer; however, this result is based on a point estimate from a single run, with no accompanying uncertainty assessment. Furthermore, the lack of detailed information regarding the baseline models—such as their capacity, layer configuration, tokenization schemes, and effective memory budgets—makes it difficult to draw meaningful structural comparisons. The "three-event cache," recurrent ConvGRU control, MPM replay window, and local-frame temporal context mentioned in the paper represent distinct implementation strategies. While the combination of these approaches suggests that "historical information" is indeed beneficial, it does not sufficiently demonstrate the existence of a unified, "world-addressed" persistence mechanism that serves as a core component across diverse settings.
  2. The reported "65.5% reduction in regret" relies on an independently supervised decision readout module, which incorporates auxiliary outcomes and a loss function conditioned on the accuracy of the historical information; consequently, this improvement cannot be attributed solely to enhancements in the "world model’s" rollouts. The decision-making evaluation is based on 80 validation states, and the regret comparisons presented in Table 1 lack uncertainty metrics. Moreover, for the MPM method, the confidence intervals for performance gains related to "volume" and "force" include zero (see Appendix E.3). In the actual study, certain information—such as action parameters and world registration—is missing. The evaluation is based on a single random seed and covers 43 targets across 20 segments. Furthermore, using "correctly ordered historical information" yields only a marginal performance improvement compared to using "shuffled historical information." While these observations are of some value, they are insufficient to fully validate the "action-conditioned architecture" and its purported advantages for planning.
Questions for authors
  1. Please provide exact baseline architectures, parameter counts, truncation rules, and per-seed paired differences for Table 1. Do the small rollout gains over History Transformer persist across training seeds?
  2. How much of the ranking gain remains with a common plug-in utility computed directly from predicted outcomes, without the separately trained decision ensemble and its history-specific supervision?
  3. Can the real-data benefit be quantified with segment-level paired intervals and acquisition-session grouping? Does a registered action-conditioned implementation outperform a comparably trained temporal-history baseline?
  4. What single state update mechanism is claimed as the main contribution? Please map each experiment to the raw-event cache, carried recurrent state, or replayed-window implementation, and clarify whether all baselines receive accumulated world-height state.

Comments on the paper

1 comment

Dear authors and reviewers,

The author-reviewer discussion period closes on November 18 (AoE). Please engage/continue to engage during this period so that we have our issues discussed thoroughly.

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.