pith. sign in

arxiv: 2607.02403 · v1 · pith:BXY6S6IPnew · submitted 2026-07-02 · 💻 cs.RO · cs.AI· cs.CV

ACID: Action Consistency via Inverse Dynamics for Planning with World Models

Gawon Seo , Dongwon Kim , Suha Kwak This is my paper

Pith reviewed 2026-07-03 11:02 UTC · model grok-4.3

classification 💻 cs.RO cs.AIcs.CV
keywords world modelsplanninginverse dynamicsaction consistencymodel-based controlembodied controldecision-time planning
0
0 comments X

The pith

ACID adds inverse-dynamics consistency to world-model planning so that predicted transitions must be recoverable from the conditioned actions.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Standard planning scores a candidate trajectory only by how close its final predicted state lies to the goal, leaving the realizability of each intermediate step unchecked. ACID inserts a cycle-consistency check: an inverse dynamics model is asked to recover the original action from each predicted transition, and the resulting residual is folded into the planning cost through a scale-invariant adaptive weight. The paper demonstrates that this term improves planning success across four different action-conditioned world models and six tasks that include rigid and deformable manipulation, articulated control, and visual navigation. The same accuracy is reached with substantially lower planning compute than the goal-only baseline.

Core claim

By requiring that the action recovered backward from a predicted state transition equals the action that was conditioned on, the cycle-consistency residual penalizes unrealizable trajectories at decision time and thereby improves both planning quality and efficiency without retraining the world model.

What carries the argument

Cycle action consistency residual: the per-step difference between the conditioned action and the action recovered by the inverse dynamics model from the predicted transition, added to the terminal goal cost via an adaptive scale-invariant weight.

If this is right

  • Planning reaches the same success rate with fewer candidate evaluations or shorter search horizons.
  • The improvement holds across rigid-body, deformable-object, and visual-navigation domains without task-specific retuning.
  • The consistency term acts as a model-error proxy that does not require additional training data or world-model updates.
  • Accuracy gains appear even when the underlying world model is already strong, indicating the check addresses a distinct failure mode.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same per-step consistency idea could be applied to other prediction modalities such as visual features or contact forces.
  • In longer-horizon problems the accumulated consistency penalty may limit compounding model drift more effectively than a terminal cost alone.
  • If the inverse dynamics model is trained on the same offline data as the world model, both may share the same distribution shift when deployed.

Load-bearing premise

The mismatch between a conditioned action and the action recovered by the inverse dynamics model serves as a reliable proxy for whether that transition will actually occur when the action is executed in the environment.

What would settle it

Run the planner on a task where the inverse dynamics model has been deliberately degraded on a known subset of transitions; if ACID then produces lower success rates than the goal-only baseline, the consistency signal is not functioning as claimed.

Figures

Figures reproduced from arXiv: 2607.02403 by Dongwon Kim, Gawon Seo, Suha Kwak.

Figure 1
Figure 1. Figure 1: Overall architecture of ACID. (Left): Decision-time planning with an action-conditioned world model: an MPC with CEM searches over candidate action sequences a0:H−1 to minimize the augmented planning cost. The current observation o0 is encoded by Eθ to z0, and the world model Fθ unrolls a latent trajectory zˆ1:H from (z0, a0:H−1). The inverse dynamics model Gϕ then takes each predicted transition (ˆzt, zˆt… view at source ↗
Figure 2
Figure 2. Figure 2: Environment suites used in our experiments. From left to right: Cube, Reacher, Push-T, Rope Manipulation, Granular Manipulation, and goal-conditioned visual navigation on RECON [34]. More details on the environments and datasets are available in Appendix C.1 [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison on the Rope and Granular. For each task, we show the real rollout in the environment and the corresponding imagined rollout from the world model under the planned action sequence. and leaves the world model untouched, so it is orthogonal to world-model improvements and can be composed with any action-conditioned world model. Qualitative results [PITH_FULL_IMAGE:figures/full_fig_p007… view at source ↗
Figure 4
Figure 4. Figure 4: Mean chamfer distance over planning steps on Granular and Rope. Lower total compute to a target quality. Since total planning compute is proportional to both the number of CEM samples and the number of planning steps, reducing either factor directly reduces the overall cost. Across both axes, Ours reaches a given quality with strictly less total planning compute than Original. On Le-WM and PLDM, the CEM-bu… view at source ↗
Figure 5
Figure 5. Figure 5: Inverse Dynamics Model(IDM) architecture. (Left): the prefix–suffix transformer that predicts the action between the latents of two consecutive frames zt and zt+1. (Right): the corresponding attention mask. Architecture. In [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Within-pool spread of the goal cost and action consistency cost. Rows are world models (Le-WM, PLDM, DINO-WM); columns show the within-pool standard deviations σg (a) and σa (b) and their ratio σg/σa (c) on a log-scaled y-axis, with one curve per task. Each point is the median over all pools at that CEM iteration. The dashed line at σg/σa = 1 marks equal within-pool spread. Why the adaptive weight works. B… view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison on Le-WM and PLDM across the OGBench-Cube, Reacher, and Push-T tasks. Le-WM is shown in the left column and PLDM in the right column. For each model and task, we show the real rollout in the environment and the corresponding imagined rollout from the world model under the planned action sequence, comparing the baseline with the baseline with ACID [PITH_FULL_IMAGE:figures/full_fig_p0… view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative results on goal-conditioned visual navigation. Comparison between ground￾truth frames and world model rollouts. The action consistency score from ACID (top) and navigation trajectories (green) are presented together. Red box highlights the step where the action consistency cost drops; at this step, the predicted next frame fails to follow the action condition. 19 [PITH_FULL_IMAGE:figures/full_… view at source ↗
read the original abstract

Decision-time planning with action-conditioned world models has become a popular paradigm for embodied control. However, the standard planning cost judges a candidate solely by how close its predicted terminal state lies to the goal, leaving the realizability of the intermediate transitions unchecked -- a predicted trajectory can look convincing while the environment rollout drifts away from it. In this paper, we propose ACID, a decision-time planning framework that introduces cycle action consistency: the action inferred backward from a predicted transition by an inverse dynamics model should recover the one that was conditioned on. We fold this per-step residual into the planning cost via a scale-invariant adaptive weight. Across four action-conditioned world models and six tasks spanning rigid and deformable manipulation, articulated control, and visual navigation, ACID consistently improves planning and matches the baseline's accuracy with substantially less planning compute.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper proposes ACID, a decision-time planning method for action-conditioned world models. It augments the standard terminal-state cost with a cycle-consistency residual ||a_t - IDM(ŝ_t, ŝ_{t+1})|| obtained from a separately trained inverse dynamics model, incorporated via a scale-invariant adaptive weight. The central claim is that this term enforces realizability of intermediate transitions and yields consistent planning improvements across four world models and six tasks (rigid/deformable manipulation, articulated control, visual navigation) while matching baseline accuracy at substantially lower planning compute.

Significance. If the consistency residual reliably proxies environment realizability rather than shared model bias, the approach would address a recognized weakness in model-based planning without extra training or interaction. The multi-model, multi-task evaluation and reported compute reduction are strengths that would make the result relevant to embodied control if substantiated.

major comments (2)
  1. [Experiments / Method (central mechanism)] The central claim that the per-step residual serves as a reliable proxy for transition realizability is load-bearing, yet the manuscript provides no direct validation (e.g., correlation between the residual and actual environment rollout error on held-out transitions). This leaves open the possibility that the reported gains arise from reduced variance in the cost rather than genuine filtering of unrealizable trajectories.
  2. [Method (cycle consistency term)] Both the action-conditioned world model and the IDM are trained on the same offline dataset; the manuscript does not analyze or mitigate the risk that systematic biases or coverage gaps are shared, causing the residual to under-penalize jointly plausible but unrealizable transitions. This directly affects whether the adaptive weighting improves realizability or merely amplifies model agreement.
minor comments (1)
  1. [Method] Notation for the adaptive weight and its scale-invariance property should be defined explicitly with an equation, as the current description leaves the precise functional form unclear.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and for highlighting the strengths of the multi-model evaluation and compute savings. We address the two major comments point-by-point below, agreeing that additional validation is warranted and proposing targeted revisions to strengthen the evidence.

read point-by-point responses

  1. Referee: [Experiments / Method (central mechanism)] The central claim that the per-step residual serves as a reliable proxy for transition realizability is load-bearing, yet the manuscript provides no direct validation (e.g., correlation between the residual and actual environment rollout error on held-out transitions). This leaves open the possibility that the reported gains arise from reduced variance in the cost rather than genuine filtering of unrealizable trajectories.

    Authors: We agree that a direct correlation analysis would provide stronger substantiation that the residual acts as a realizability proxy rather than simply reducing cost variance. In the revised manuscript we will add an experiment on held-out transitions that reports the Pearson correlation between the per-step ACID residual and the actual environment rollout error (state discrepancy after executing the planned action sequence). This analysis will appear in a new subsection of the experiments or an appendix. revision: yes

  2. Referee: [Method (cycle consistency term)] Both the action-conditioned world model and the IDM are trained on the same offline dataset; the manuscript does not analyze or mitigate the risk that systematic biases or coverage gaps are shared, causing the residual to under-penalize jointly plausible but unrealizable transitions. This directly affects whether the adaptive weighting improves realizability or merely amplifies model agreement.

    Authors: The shared-data concern is legitimate. While the forward and inverse objectives differ and can produce distinct error patterns, we will add an explicit discussion and supporting analysis in the revision: we will quantify disagreement between the world model and IDM on held-out data and show how the adaptive weight modulates the residual in those cases. We will also state the limitation clearly and note that training the IDM on a disjoint dataset (when available) is a natural future direction. These changes clarify assumptions without altering the method. revision: partial

Circularity Check

0 steps flagged

No significant circularity; consistency residual is an independent check

full rationale

The ACID framework trains an action-conditioned world model and a separate inverse dynamics model on the same offline dataset, then uses the per-step residual between a conditioned action and the IDM-recovered action as an additive term in the planning cost. This residual is computed from an external model rather than being defined in terms of the world model's own predictions or fitted parameters. No equations reduce the consistency term to a self-referential quantity by construction, no uniqueness theorems are imported from prior self-citations, and no ansatz is smuggled via citation. The reported gains are empirical across multiple models and tasks, making the derivation self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The approach rests on the assumption that a reliable inverse dynamics model exists and that its residuals meaningfully indicate planning quality. No new physical entities are postulated.

free parameters (1)
  • scale-invariant adaptive weight
    Used to incorporate the consistency residual into the planning cost; its computation or fitting details are not specified in the abstract.
axioms (1)
  • domain assumption An inverse dynamics model can accurately recover the conditioned action from a predicted state transition
    This underpins the cycle consistency check described in the abstract.

pith-pipeline@v0.9.1-grok · 5668 in / 1214 out tokens · 59754 ms · 2026-07-03T11:02:54.171997+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

40 extracted references · 13 canonical work pages · 10 internal anchors

  1. [1]

    World Models

    D. Ha and J. Schmidhuber. World models.arXiv preprint arXiv:1803.10122, 2(3), 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  2. [2]

    A. Bar, G. Zhou, D. Tran, T. Darrell, and Y . LeCun. Navigation world models. InProc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2025

    2025

  3. [3]

    D. Kim, G. Seo, J. Lee, M. Cho, and S. Kwak. Planning in 8 tokens: A compact discrete tokenizer for latent world model. InProc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2026

    2026

  4. [4]

    G. Zhou, H. Pan, Y . LeCun, and L. Pinto. Dino-wm: World models on pre-trained visual features enable zero-shot planning. InProc. International Conference on Machine Learning (ICML), 2025

    2025

  5. [5]

    L. Maes, Q. L. Lidec, D. Scieur, Y . LeCun, and R. Balestriero. Leworldmodel: Stable end-to-end joint-embedding predictive architecture from pixels.arXiv preprint arXiv:2603.19312, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  6. [6]

    V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning

    M. Assran, A. Bardes, D. Fan, Q. Garrido, R. Howes, M. Muckley, A. Rizvi, C. Roberts, K. Sinha, A. Zholus, et al. V-jepa 2: Self-supervised video models enable understanding, prediction and planning.arXiv preprint arXiv:2506.09985, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  7. [7]

    Sobal, W

    U. Sobal, W. Zhang, K. Cho, R. Balestriero, T. G. Rudner, and Y . LeCun. Learning from reward-free offline data: A case for planning with latent dynamics models.Advances in Neural Information Processing Systems, 38:43905–43941, 2026

    2026

  8. [8]

    H. Nam, Q. L. Lidec, L. Maes, Y . LeCun, and R. Balestriero. Causal-jepa: Learning world models through object-level latent interventions. InProc. International Conference on Machine Learning (ICML), 2026

    2026

  9. [9]

    Hansen, H

    N. Hansen, H. Su, and X. Wang. Td-mpc2: Scalable, robust world models for continuous control. InProc. International Conference on Learning Representations (ICLR), 2024

    2024

  10. [10]

    Morari and J

    M. Morari and J. H. Lee. Model predictive control: past, present and future.Computers & Chemical Engineering, 23:667–682, 1999

    1999

  11. [11]

    De Boer, D

    P.-T. De Boer, D. P. Kroese, S. Mannor, and R. Y . Rubinstein. A tutorial on the cross-entropy method.Annals of operations research, 134(1):19–67, 2005

    2005

  12. [12]

    B. Chen, D. Mart´ı Mons´o, Y . Du, M. Simchowitz, R. Tedrake, and V . Sitzmann. Diffusion forcing: Next-token prediction meets full-sequence diffusion. InProc. Neural Information Processing Systems (NeurIPS), 2024

    2024

  13. [13]

    Y . Du, M. Yang, P. Florence, F. Xia, A. Wahid, B. Ichter, P. Sermanet, T. Yu, P. Abbeel, J. B. Tenenbaum, et al. Video language planning. InProc. International Conference on Learning Representations (ICLR), 2024

    2024

  14. [14]

    A. Hu, L. Russell, H. Yeo, Z. Murez, G. Fedoseev, A. Kendall, J. Shotton, and G. Corrado. Gaia-1: A generative world model for autonomous driving.arXiv preprint arXiv:2309.17080, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  15. [15]

    S. Gao, J. Yang, L. Chen, K. Chitta, Y . Qiu, A. Geiger, J. Zhang, and H. Li. Vista: A generalizable driving world model with high fidelity and versatile controllability. InProc. Neural Information Processing Systems (NeurIPS), 2024

    2024

  16. [16]

    S. Ye, Y . Ge, K. Zheng, S. Gao, S. Yu, G. Kurian, S. Indupuru, Y . L. Tan, C. Zhu, J. Xiang, et al. World action models are zero-shot policies.arXiv preprint arXiv:2602.15922, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  17. [17]

    Bruce, M

    J. Bruce, M. D. Dennis, A. Edwards, J. Parker-Holder, Y . Shi, E. Hughes, M. Lai, A. Mavalankar, R. Steigerwald, C. Apps, et al. Genie: Generative interactive environments. InProc. Interna- tional Conference on Machine Learning (ICML), 2024. 9

    2024

  18. [18]

    Y . Du, S. Yang, B. Dai, H. Dai, O. Nachum, J. Tenenbaum, D. Schuurmans, and P. Abbeel. Learning universal policies via text-guided video generation. InProc. Neural Information Processing Systems (NeurIPS), 2023

    2023

  19. [19]

    S. Kim, S. Lee, and M. Cho. Freeaction: Training-free techniques for enhanced fidelity of trajectory-to-video generation.arXiv preprint arXiv:2509.24241, 2025

    work page arXiv 2025

  20. [20]

    K. Song, B. Chen, M. Simchowitz, Y . Du, R. Tedrake, and V . Sitzmann. History-guided video diffusion. InProc. International Conference on Machine Learning (ICML), 2025

    2025

  21. [21]

    D. Kim, N. Kim, and S. Kwak. Improving cross-modal retrieval with set of diverse embeddings. InProc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023

    2023

  22. [22]

    Song and M

    Y . Song and M. Soleymani. Polysemous visual-semantic embedding for cross-modal retrieval. InProc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019

    2019

  23. [23]

    S. Chun, S. J. Oh, R. S. De Rezende, Y . Kalantidis, and D. Larlus. Probabilistic embeddings for cross-modal retrieval. InProc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021

    2021

  24. [24]

    Rigter, T

    M. Rigter, T. Gupta, A. Hilmkil, and C. Ma. A VID: Adapting video diffusion models to world models. InReinforcement Learning Conference, 2025

    2025

  25. [25]

    A. Xie, O. Rybkin, D. Sadigh, and C. Finn. Latent diffusion planning for imitation learning. In Proc. International Conference on Machine Learning (ICML), 2025

    2025

  26. [26]

    A. Ye, B. Wang, C. Ni, G. Huang, G. Zhao, H. Li, H. Li, J. Li, J. Lv, J. Liu, et al. Gigaworld- policy: An efficient action-centered world–action model.arXiv preprint arXiv:2603.17240, 2026

    work page arXiv 2026

  27. [27]

    C. Zhu, R. Yu, S. Feng, B. Burchfiel, P. Shah, and A. Gupta. Unified world models: Coupling video and action diffusion for pretraining on large robotic datasets. InRobotics: Science and Systems (RSS), 2025

    2025

  28. [28]

    Brandfonbrener, O

    D. Brandfonbrener, O. Nachum, and J. Bruna. Inverse dynamics pretraining learns good representations for multitask imitation. InProc. Neural Information Processing Systems (NeurIPS), 2023

    2023

  29. [29]

    Baker, I

    B. Baker, I. Akkaya, P. Zhokov, J. Huizinga, J. Tang, A. Ecoffet, B. Houghton, R. Sampedro, and J. Clune. Video pretraining (vpt): Learning to act by watching unlabeled online videos. In Proc. Neural Information Processing Systems (NeurIPS), 2022

    2022

  30. [30]

    J. Jang, S. Ye, Z. Lin, J. Xiang, J. Bjorck, Y . Fang, F. Hu, S. Huang, K. Kundalia, Y .-C. Lin, et al. Dreamgen: Unlocking generalization in robot learning through video world models.arXiv preprint arXiv:2505.12705, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  31. [31]

    Lipman, R

    Y . Lipman, R. T. Chen, H. Ben-Hamu, M. Nickel, and M. Le. Flow matching for generative modeling. InProc. International Conference on Learning Representations (ICLR), 2023

    2023

  32. [32]

    Q. Liu. Rectified flow: A marginal preserving approach to optimal transport.arXiv preprint arXiv:2209.14577, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  33. [33]

    $\pi_0$: A Vision-Language-Action Flow Model for General Robot Control

    K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, B. Ichter, et al. π0: A vision-language-action flow model for general robot control.arXiv preprint arXiv:2410.24164, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  34. [34]

    D. Shah, B. Eysenbach, N. Rhinehart, and S. Levine. Rapid exploration for open-world navigation with latent goal models. InAnnual Conference on Robot Learning (CoRL), 2021. 10

    2021

  35. [35]

    Sturm, W

    J. Sturm, W. Burgard, and D. Cremers. Evaluating egomotion and structure-from-motion approaches using the tum rgb-d benchmark. InProc. of the Workshop on Color-Depth Camera Fusion in Robotics at the IEEE/RJS International Conference on Intelligent Robot Systems (IROS), 2012

    2012

  36. [36]

    Zhang, P

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang. The unreasonable effectiveness of deep features as a perceptual metric. InProc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018

    2018

  37. [37]

    Loshchilov and F

    I. Loshchilov and F. Hutter. Decoupled weight decay regularization. InProc. International Conference on Learning Representations (ICLR), 2019

    2019

  38. [38]

    LeJEPA: Provable and Scalable Self-Supervised Learning Without the Heuristics

    R. Balestriero and Y . LeCun. Lejepa: Provable and scalable self-supervised learning without the heuristics.arXiv preprint arXiv:2511.08544, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  39. [39]

    DeepMind Control Suite

    Y . Tassa, Y . Doron, A. Muldal, T. Erez, Y . Li, D. d. L. Casas, D. Budden, A. Abdolmaleki, J. Merel, A. Lefrancq, et al. Deepmind control suite.arXiv preprint arXiv:1801.00690, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  40. [40]

    Zhang, B

    K. Zhang, B. Li, K. Hauser, and Y . Li. Adaptigraph: Material-adaptive graph-based neural dynamics for robotic manipulation.arXiv preprint arXiv:2407.07889, 2024. 11 Appendix A Implementation Details of IDM Figure 5:Inverse Dynamics Model(IDM) architecture.(Left): the prefix–suffix transformer that predicts the action between the latents of two consecutiv...

    work page arXiv 2024