acceptodds
Under review as a conference paper at ICLR 2027

ExAbs: Learning Execution-Aware Abstractions for Hierarchical World Models

Abstract

Hierarchical world models extend planning horizons by letting a high-level model propose waypoints for a short-horizon controller to reach. Their temporal abstractions, however, are typically learned from offline trajectories without regard to the controller that must realize them: a transition that is easy to predict may be infeasible for the controller within its budget, and similarity between full latent states can overlook the specific change a waypoint requests. We propose execution-aware abstraction learning (ExAbs), in which the controller shapes both what the hierarchy learns to request and how each request is pursued. For the former, an Achievability Model (AM) learns from closed-loop execution the probability that the controller completes a waypoint within a given budget; AM selects the segment boundaries that train the high-level model and scores candidate waypoints during planning. For the latter, Planner-Executor Transition Abstraction Learning (PETAL) distills the decisions of a frozen planner into compact, source-relative effect representations, so that local action selection compares requested and predicted changes rather than endpoint proximity alone. Collecting AM labels with the PETAL-enabled controller ties the learned temporal units to the executor that carries them out. Across seven simulated tasks on four world-model backbones and three physical-robot tasks, ExAbs improves success over the hierarchical baseline on every task without lengthening the local planning horizon, e.g., from 77.5% to 90.0% on hard Maze queries and from 0% to 21.7% on a sequential Button–Window task, while also shortening successful executions.

Primary area:
applications to robotics, autonomy, planning
Keywords:
hierarchical world models, temporal abstraction, model-based planning, latent planning, achievability, executability, robot learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_X9n9OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes ExAbs, which is a method for learning execution-aware temporal abstractions in hierarchical world models. Standard hierarchical models learn high-level waypoints from offline trajectories without considering whether the low-level controller can actually reach them within budget. ExAbs addresses this with two components: an Achievability Model (AM) that learns from closed-loop execution the probability a waypoint is reachable within a given budget and uses this to select segment boundaries and score candidate waypoints during planning; and PETAL, which distills a frozen planner’s decisions into compact source-relative effect representations so local action selection compares requested versus predicted changes rather than endpoint proximity alone. Collecting AM labels with the PETAL-enabled controller ties learned abstractions to the actual executor. Across seven simulated tasks on four world-model backbones and three real-robot tasks,

Critical strengths

1. AM innovatively uses achievability to judge the quality of candidate transitions: it asks whether the controller can actually complete a transition within a given budget, which rules out the "predictable but infeasible" subgoals that arise in conventional approaches.

2. AM and PETAL are not isolated modules; the authors organise them into a coherent whole. The implementation is careful as well: PETAL is frozen before AM labels are collected, and AM is frozen before the boundaries are selected, which avoids the problem that a changed local objective invalidates previously collected labels. I particularly appreciate this.

3. The paper is well written.

Critical weaknesses

1. PETAL's motivation is unclear. The authors state that "full-state similarity is shaped largely by content that the actions do not change", and §5.2 / Appendix K.2 further gloss this as "appearance changes that leave the physical transition unchanged". Unless I misunderstand the setup, every experiment in the paper uses a fixed-view camera, and on the physical robot even the scene fixtures remain fixed. Appearance should therefore not change within an episode; what changes is the part related to the agent's pose. Moreover, since the requested waypoint ww and the candidate endpoint z^i\hat z_i both derive from the same ztz_t, the content they share cancels in the difference z^i−w\hat z_i-w. What can actually differentiate candidates is then only candidate-correlated prediction error, together with the mismatch between the retrained high-level predictor FHF_H (which produces ww) and the frozen low-level predictor FLF_L (which produces z^i\hat z_i). It is not the appearance that the authors emphasise.

2. PETAL does contribute, but no internal ablation is provided, and the distribution of the gains points to another mechanism. The improvement of +PETAL over L2 is real (Table 5: Reach-Wall 36.5→47.9, Counterflow 37.5→56.5, Two-room 9.4→15.6; Table S11: PushT d=50d=50, 28.7→32.7), yet as far as I can see the paper does not ablate PETAL's internal design. The distribution of the gains also points to a different mechanism: +PETAL is largest exactly on the tasks where the agent must temporarily move away from the waypoint (Reach-Wall +11.5, Counterflow +19.0, Two-room +6.3) and small where it need not (Maze-Hard +2.5; PushT +2 to +4). This is what §3.3 describes when it says that the anchored terminal comparison replaces the original path-summed goal cost: a path-summed cost toward a single waypoint penalises precisely those necessary detours, whereas appearance invariance would not concentrate the gains on obstacle and constraint tasks. I therefore suspect that what actually drives PETAL is not the part the authors advertise.

Questions for authors

1. What is the observation setup for the simulated tasks: a single world-fixed camera, or an egocentric view? Is the camera pose constant within an episode? Are goal images rendered under the same camera, lighting and materials?

2. Which part of PETAL produces the gain, and does the paper provide the corresponding ablation (for example ω=0\omega=0, i.e. the anchor alone; a fixed metric instead of the learned one; the support mask disabled; the top-128 shortlist removed; and terminal versus path-summed comparison)?

7 comments
AuthorsOpenReview ↗

[Part 1 of 3 — Q1 & W1: Observation Setup & Static-Background Analysis]

Dear Reviewer X9n9,

Thank you for the careful reading and for the precise questions about PETAL. Part 1 addresses Q1 and W1, Part 2 covers endpoint scoring (T0/T1) and the learned effect encoder (A/R/P), and Part 3 reports the remaining internal ablations (FixedD, Uniform, NoSL). All updates appear in the uploaded revision.

[Q1 & W1] Camera Setup and Why Static Backgrounds Do Not Cancel Exactly

1. Answer to Q1 (Appendix H.1). All seven primary simulated tasks (plus Button–Drawer) use a single world-fixed camera rather than an egocentric view, with constant camera pose within and across episodes: global 2D renderings in Counterflow and Two-room, fixed top-down views in Maze and PushT, and fixed external views in Cube, Reach-Wall, Button–Window, and Button–Drawer (supplementary Visual AntMaze follows OGBench's default agent-tracking third-person camera). Goal images are rendered by the same simulator with the exact same camera, lighting, and materials as episode observations, while intermediate hierarchical waypoints are model-predicted latent states.

2. Response to W1 (Additive cancellation vs. nonlinear entanglement). We agree with your decomposition: if the encoder EE mapped static background to an additive latent offset shared across candidates, that offset would cancel in z^i−w\hat{z}_i - w, leaving only (a) candidate-correlated FL\mathcal F_L prediction error and (b) FH\mathcal F_H-vs.-FL\mathcal F_L horizon/predictor mismatch. In practice, a third term enters: (c) learned ViT and ConvNet encoders are nonlinear and entangle static background context with agent and object pose, so background is not an additive constant even when camera and scene appearance stay fixed.

We test (c) directly with Control 1 (Real outcomes) and examine stress-test robustness with Controls 2–4 (Section 5.2, Appendix K.4, Table S16), holding camera geometry, physical starts/goals, dynamics, models, and budgets fixed with online AM disabled:

  • Control 1 isolates (c) with zero prediction error: We fix candidate action sequences and render source, real target, and each candidate's ground-truth physical outcome under one shared background, then rescore after changing that shared background. Neither (a) FL\mathcal F_L error nor (b) FH\mathcal F_H/FL\mathcal F_L mismatch enters. If shared backgrounds cancelled, candidate selection would never change. Instead, base-scoring attainment drops in three of the five tasks (Cube, Two-room, PushT), and PETAL reduces the drop in each of those three:
Task (Requests)Ctrl 1: Real Outcomes — BaseCtrl 1: Real Outcomes — PETALCtrl 2: Full Episode — w/o PETALCtrl 2: Full Episode — w/ PETAL
Counterflow (200)82.0 →\to 83.093.0 →\to 96.540.7 →\to 19.260.2 →\to 50.8
Cube d=50d=50 (50)78.0 →\to 70.078.0 →\to 74.057.3 →\to 36.756.7 →\to 42.0
Two-room med. (32)37.5 →\to 15.640.6 →\to 25.015.6 →\to 12.531.3 →\to 28.1
PushT d=50d=50 (50)56.0 →\to 44.054.0 →\to 48.028.0 →\to 20.032.0 →\to 26.0
Button–Window (60)10.0 →\to 10.010.0 →\to 8.31.7 →\to 1.718.3 →\to 15.0
  • Control 2 (Full closed-loop execution): Current, historical, and goal images share one changed background, kept static throughout the episode. We report full-task success averaged over three seeds on Counterflow, Cube, and Reach-Wall, and one elsewhere. PETAL reduces degradation in several settings, including Reach-Wall (47.9$\to$40.6% vs. 31.9$\to$20.5% without PETAL).
  • Control 3 (Expert local targets): We fix the physical start and a local target from an expert trajectory. Only the goal image's background changes; the execution environment is unchanged. We measure local-target attainment.
  • Control 4 (Handoff local targets): We repeat Control 3 with the physical state reached at the first hierarchical handoff as the target, extending the test to goals produced during hierarchical execution.
TaskCtrl 3: Expert Target (w/o ∣\mid w/ PETAL)Ctrl 4: Handoff Target (w/o ∣\mid w/ PETAL)
Counterflow100$\to$62.0 ∣\mid 99.5$\to$97.0100$\to$55.5 ∣\mid 99.5$\to$97.0
Cube d=50d=5098.0$\to$54.0 ∣\mid 98.0$\to$78.098.0$\to$58.0 ∣\mid 98.0$\to$80.0
Two-room med.96.9$\to$62.5 ∣\mid 93.8$\to$78.1100$\to$75.0 ∣\mid 100$\to$81.3
PushT d=50d=5074.0$\to$52.0 ∣\mid 76.0$\to$64.080.0$\to$60.0 ∣\mid 82.0$\to$72.0
Button–Window33.3$\to$20.0 ∣\mid 48.3$\to$51.710.0$\to$6.7 ∣\mid 16.7$\to$16.7

Across these five tasks, PETAL improves attainment under the changed goal background in both controls, suggesting benefits for local execution toward both expert and handoff targets. Meanwhile, native unshifted closed-loop gains are isolated directly in Part 2 below.

AuthorsOpenReview ↗

[Part 2 of 3 — Q2 & W2: Endpoint Cost vs. Learned Effect Representation]

[Q2 & W2 — Part A] Endpoint Scoring (T0/T1) vs. Learned Effect Encoder (A/R/P)

Thank you for the concrete ablation list. We ran all requested controls under matched world models, queries, budgets, and planning seeds with online AM disabled throughout (Section 5.3, Appendix L.4, Tables S17–S19). In short: switching from path-aggregated to endpoint-only base cost changes success by −3.1-3.1 to +1.2+1.2 pp, whereas holding the anchor fixed and adding the learned effect encoder Φ\Phi adds +2.0+2.0 to +15.0+15.0 pp across all seven tasks (+9.1 pp pooled over 560 episodes).

1. Path-aggregated vs. endpoint-only base cost without PETAL (T0 vs. T1, Table S17). Reach-Wall, PushT, Cube, Button–Window, and Button–Drawer already evaluate endpoint error in their native cost or anchor (Section 3.3, Appendix D.2), so replacing path aggregation cannot explain +PETAL's +11.5 pp gain on Reach-Wall. Path aggregation is replaced only in Maze (path sum), Two-room (path mean), and Counterflow (final 5-step mean). Testing native path cost (T0) against endpoint-only distance (T1) without PETAL never improves success by more than +1.2 pp and lowers the other three settings:

TaskQueriesT0 (Native Path Cost)T1 (Endpoint-Only Cost)T1 −- T0 (pp)
Counterflow20040.67%41.83%+1.17
Two-room6410.94%7.81%-3.13
Maze Medium4095.00%92.50%-2.50
Maze Hard4082.50%80.00%-2.50

2. Anchor Alone vs. Random Encoder vs. Learned PETAL (A/R/P, Table S18). Holding the complete PETAL anchor (including its temporal and spatial form) and single-seed evaluation protocol fixed across arms, we compare:

  • A (ω=0\omega=0): PETAL's anchor alone, with the effect branch disabled.
  • R: Identically structured random encoder frozen at initialization, with its comparison network retrained from scratch on the same teacher data, budget, and validation rule (isolating Eq. 5 + Eq. D4–D5 supervision of Φ\Phi).
  • P: Trained PETAL encoder and comparison network.
TaskQueriesA (ω=0\omega=0, Anchor)R (Random Φ\Phi)P (Learned PETAL)P −- A (pp)P −- R (pp)
Reach-Wall9636.46%42.71%47.92%+11.46+5.21
Counterflow20042.50%40.50%54.50%+12.00+14.00
Button–Window603.33%0.00%18.33%+15.00+18.33
Two-room6410.94%4.69%15.63%+4.69+10.94
PushT d=50d=505028.00%22.00%32.00%+4.00+10.00
Maze Hard4085.00%80.00%87.50%+2.50+7.50
Cube d=50d=505060.00%56.00%62.00%+2.00+6.00
Total (560)56036.96% (207)35.00% (196)46.07% (258)+9.11+11.07

With the anchor held fixed, learned effects (P −- A) add +11.5 pp on Reach-Wall, +12.0 pp on Counterflow, +15.0 pp on Button–Window, and +4.7 pp on Two-room. P also exceeds R in all 7 tasks (+5.2 to +18.3 pp), showing that the gain comes from learning Φ\Phi rather than adding a trainable scorer. Obstacle and precondition tasks benefit most because over a short 10–20 step local horizon, candidates stalling near a barrier and candidates initiating a tangential detour or button press can have similar endpoint distance AiA_i, whereas directional source-relative effects separate clearance motion from stalled rollouts.

AuthorsOpenReview ↗

[Part 3 of 3 — Q2 & W2: Metric, Spatial Support & Shortlist Ablations]

[Q2 & W2 — Part B] Remaining Design Choices in Table S19

Below we examine three further design choices in Appendix L.4 and Table S19 under matched evaluation protocols with online AM disabled throughout. Each ablation covers every setting whose adapter includes that component.

FixedD retains the learned effect encoder and replaces the learned comparison metric with squared Euclidean distance, testing the contribution of metric learning. Uniform replaces selective spatial support κ∗\kappa^* with uniform spatial weights, testing the contribution of spatial selection. NoSL scores the full candidate pool instead of the top-128 shortlist, testing whether candidate filtering materially affects success.

3. Learned Metric vs. Squared Euclidean Distance Across All 9 Settings (Table S19a):

TaskFixedD (Fixed L2L_2)P (Learned Metric)P −- FixedD (pp)
Maze Medium92.50%95.00%+2.50
Maze Hard85.00%87.50%+2.50
PushT d=50d=5028.00%30.00%+2.00
Two-room14.06%15.63%+1.56
Reach-Wall46.88%47.92%+1.04
Counterflow56.17%56.17%0.00
PushT d=75d=7518.67%18.67%0.00
Button–Window18.33%18.33%0.00
Cube d=50d=5064.00%62.00%-2.00

4. Selective Spatial Support vs. Uniform Pooling Across All 6 Spatial-Grid Settings (Table S19b): LeWM uses 1D vector latents on PushT and Cube without a spatial feature grid. Across all six tasks with spatial feature maps, selective support κ∗\kappa^* is strictly higher in 5 settings and ties at 95.0% on Maze Medium:

Spatial-Grid TaskUniform (No κ∗\kappa^*)P (Selective κ∗\kappa^*)P −- Uniform (pp)
Reach-Wall44.79%47.92%+3.13
Counterflow53.67%56.17%+2.50
Maze Hard85.00%87.50%+2.50
Button–Window16.67%18.33%+1.67
Two-room14.06%15.63%+1.56
Maze Medium95.00%95.00%0.00

5. Top-128 Candidate Shortlist vs. Full-Pool Scoring Across All 4 Shortlisted Settings (Table S19c): The remaining adapters already score all candidates without truncation. Removing the top-128 shortlist changes success by −1.7-1.7 to +2.5+2.5 pp, indicating it serves as a compute optimization rather than the source of PETAL's gains:

Shortlisted TaskNoSL (Full Pool)P (Top-128 Shortlist)P −- NoSL (pp)
Maze Medium92.50%95.00%+2.50
Maze Hard87.50%87.50%0.00
Counterflow57.33%56.17%-1.17
Button–Window20.00%18.33%-1.67

These ablations show that metric learning and selective spatial support can improve task success, while the shortlist has small effects that vary across tasks. The learned metric matches or exceeds FixedD in 8 of 9 settings (strictly higher in 5), and selective spatial support κ∗\kappa^* matches or exceeds Uniform in all 6 spatial settings (strictly higher in 5). Since the success-rate changes from these three components are limited to about 3.1 pp, they can be viewed as secondary design choices. Taken together, Tables S17–S19 suggest that PETAL's gains arise primarily from learning the source-relative effect representation Φ\Phi.

Reviewer_aeT3OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Strength:

  • The mismatch between predictive abstraction and executable control is a good identification.
  • The evaluation spans diverse tasks and predictive backbones.
  • The paper provides useful component and representation analyses.

Weakness:

  • Most of the environments studied in this work is limited in scale. It remains unclear how the approach work for a scalably real robotic long-horizon task, e.g., in previous works [1,2,3].
  • The PETAL ablations do not isolate the benefit of the learned effect representation.
  • The evidence for reliable physical improvement remains limited.

[1] Wang, Qianwei, et al. "Unifying deep predicate invention with pre-trained foundation models." IEEE Transactions on Robotics (2026). [2] Hafner, Danijar, et al. "Deep hierarchical planning from pixels." Advances in Neural Information Processing Systems 35 (2022): 26091-26104. [3] Liang, Yichao, et al. "Visualpredicator: Learning abstract world models with neuro-symbolic predicates for robot planning." International Conference on Learning Representations. Vol. 2025. 2025.

Critical strengths

See summary

Critical weaknesses

See summary

Questions for authors

See summary

3 comments
AuthorsOpenReview ↗

[Part 1 of 2 — W1 & W2.1: Long-Horizon Tasks & Learned vs. Random Effect Encoder]

Dear Reviewer aeT3,

Thank you for your supportive evaluation and constructive suggestions. Below we address all three weaknesses (W1–W3) with two new long-horizon benchmarks, controlled internal ablations of PETAL, and a completed repeat evaluation of Basket Transfer. Part 1 addresses task scale (W1) and isolating PETAL's learned effect encoder (W2.1); Part 2 covers spatial support (W2.2) and physical-robot repeatability (W3). All new experiments and discussions appear in the uploaded revision.

[W1] Task Scale and Long-Horizon Robotic Execution

Following your suggestion and the task families in UniPred (Wang et al., 2026), Director (Hafner et al., 2022), and VisualPredicator (Liang et al., 2025), we expanded Sections 1–2 and added Button–Drawer and Visual AntMaze alongside Button–Window in Section 4.2, Appendix H.4, and Table S15:

  1. Dependent manipulation stages: UniPred and VisualPredicator enforce action preconditions via symbolic predicates. Button–Window and Button–Drawer impose the same precondition dependency directly from pixels: the window or drawer moves only after the button is pressed, ExAbs receives only the final goal image without stage labels, and pressing the button alone does not count. Successful Button–Window executions average ~210 environment steps, more than 10×10\times the 20-step local planning horizon.
  2. Articulated quadruped navigation: On OGBench visual-antmaze-medium from the AntMaze family in Director, an 8-DoF torque-driven quadruped coordinates continuous joint control while routing around maze walls. Using the official 50-task set, frozen low-level backbone, 350-step budget, and planner configuration from concurrent H-JEPA (Zhang et al., 2026), ExAbs adds AM segmentation, online AM scoring, and PETAL to the shared-state HWM baseline.
TaskKey RequirementHWMExAbs (Ours)Gain
Button–WindowPress button →\to slide window0/60 (0.0%)13/60 (21.7%)+21.7 pp
Button–Drawer (New)Press button →\to open drawer0/60 (0.0%)11/60 (18.3%)+18.3 pp
Visual AntMaze (New)8-DoF quadruped + maze routing18/50 (36.0%)23/50 (46.0%)+10.0 pp

HWM completes neither two-stage dependency chain, whereas ExAbs completes both and raises Visual AntMaze success by +10.0 pp on the same backbone.

[W2.1] Learned vs. Random Effect Encoder in PETAL (Table S18)

Table 5 toggles PETAL as a whole, so it does not separate the learned effect encoder from the anchor and scoring pipeline. In Section 5.3 and Table S18, we hold world models, high-level planner, queries, planning seed, and budget fixed, with online AM disabled across all arms so variants differ only in local scoring:

  • A (Anchor only): PETAL's anchor alone (ω=0\omega=0), with the effect branch disabled.
  • R (Random encoder): Identical encoder architecture frozen at random initialization, with its comparison network retrained from scratch on the same teacher data, budget, and validation rule. Thus P minus R isolates the benefit of learning the encoder itself.
  • P (Learned PETAL): Trained effect encoder and comparison network.
TaskA (Anchor)R (Random Φ\Phi)P (PETAL)P −- R (pp)P −- A (pp)
Reach-Wall35/96 (36.5%)41/96 (42.7%)46/96 (47.9%)+5.2+11.5
Counterflow85/200 (42.5%)81/200 (40.5%)109/200 (54.5%)+14.0+12.0
Button–Window2/60 (3.3%)0/60 (0.0%)11/60 (18.3%)+18.3+15.0
Two-room7/64 (10.9%)3/64 (4.7%)10/64 (15.6%)+10.9+4.7
PushT d=50d=5014/50 (28.0%)11/50 (22.0%)16/50 (32.0%)+10.0+4.0
Maze Hard34/40 (85.0%)32/40 (80.0%)35/40 (87.5%)+7.5+2.5
Cube d=50d=5030/50 (60.0%)28/50 (56.0%)31/50 (62.0%)+6.0+2.0
Total (560 episodes)207/560 (37.0%)196/560 (35.0%)258/560 (46.1%)+11.1+9.1

P outperforms R in all 7 tasks (+5.2 to +18.3 pp; +11.1 pp pooled over 560 episodes) and outperforms A in all 7 tasks (+2.0 to +15.0 pp; +9.1 pp pooled). R falls below A in 6 of 7 tasks because unlearned random features inject noise into candidate comparison, so the effect branch helps only when its representation is learned.

AuthorsOpenReview ↗

[Part 2 of 2 — W2.2 & W3: Spatial Support & Physical-Robot Repeatability]

[W2.2] Selective vs. Uniform Spatial Support (Table S19b)

This ablation tests whether selecting spatial regions for comparison is more helpful than uniformly scoring the whole image when using learned features. We keep the trained encoder, comparison network, world models, task queries, planning budget, and planning seeds fixed within each comparison. Uniform replaces the selective support κ∗\kappa^* with uniform spatial weights. Across all six spatial-grid settings, this replacement lowers success in 5 of 6 settings and leaves it unchanged at 95.0% on Maze Medium:

Spatial-Grid TaskUniform (No κ∗\kappa^*)P (Selective κ∗\kappa^*)Gain (pp)
Reach-Wall44.8%47.9%+3.1
Counterflow53.7%56.2%+2.5
Maze Hard85.0%87.5%+2.5
Button–Window16.7%18.3%+1.7
Two-room14.1%15.6%+1.6
Maze Medium95.0%95.0%0.0

Learning the effect representation Φ\Phi thus drives the primary gain across every task, while selective support κ∗\kappa^* provides a smaller complementary gain (+1.6 to +3.1 pp).

[W3] Repeatability and Scope of Physical-Robot Improvements

To test hardware run-to-run repeatability, we ran an independent 50-trial evaluation of Basket Transfer with the frozen ExAbs model, held-out configurations, and interleaved resets documented in Section 4.4 and Appendix H.1. Below we report both the Basket Transfer repeat and pooled totals across all 150 matched trials of the three-task physical benchmark in Table 4:

1. Basket Transfer Repeatability (50 trials per run):

Method & CohortSuccesses / TrialsSuccess Rate
Flat V-JEPA 2-AC (Table 4)2 / 504.0%
HWM (Table 4)10 / 5020.0%
ExAbs — Original evaluation (Table 4)14 / 5028.0%
ExAbs — Independent 50-trial repeat (New)15 / 5030.0%

2. Pooled 3-Task Physical Benchmark (Basket, Drawer, Raised Transfer; 150 matched trials per method):

MethodTotal Successes / TrialsOverall Success Rate
Flat V-JEPA 2-AC2 / 1501.3%
HWM27 / 15018.0%
ExAbs (Ours)37 / 15024.7%

The independent repeat on Basket Transfer (30.0%) matches the original evaluation (28.0%) within 1 trial and remains 8–10 pp above HWM, and pooled success across the 150 matched trials increases from 18.0% to 24.7%. Due to hardware time constraints during the rebuttal window, we could not repeat Drawer Loading and Raised Transfer, and we note in Section 6 that larger hardware sample sizes and multi-view sensing remain important next steps.

Regarding the comparison with UniPred [1], Director [2], VisualPredicator [3], and DayDreamer [5]: those systems execute learned low-level policies or grounded skill operators, whereas our setup follows V-JEPA 2-AC [6] and HWM by optimizing continuous actions online from a single uncalibrated RGB view. In qualitative rollouts, ExAbs orders intermediate subgoals across grasp, obstacle clearance, and placement. We believe that single-camera depth ambiguity during fine grasp closure or drawer engagement may contribute to some of the observed failures (Section 4.4, Appendix H.1).

References: [1] Wang et al., IEEE T-RO 2026. [2] Hafner et al., NeurIPS 2022. [3] Liang et al., ICLR 2025. [4] Zhang et al., H-JEPA, arXiv:2610.06805, 2026. [5] Wu et al., CoRL 2022. [6] Assran et al., V-JEPA 2, arXiv:2506.09985, 2025.

AuthorsOpenReview ↗

Thank You for Your Feedback & Final Follow-up

Dear Reviewer aeT3,

We have responded to your comments and updated the paper accordingly. As the discussion period is nearing its end, we would be happy to answer any further questions or discuss any remaining concerns. Thank you for your time and help!

Best regards,
The Authors

Reviewer_8NtZOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes a pipeline to learn a hierarchical world model that is effective for planning. The proposed approach is based on providing supervision to align the requirements needed from each level of the hierarchy: the higher-level must propose achievable goals, the low-level must be accurate enough to achieve the higher-level commands. The approach is scaffolded, contrary to the end-to-end approach methods that try to learn the hierarchy directly from the dataset. They provide a good empirical evaluation that showcases the effectiveness of their hierarchy in a variety of simulated planning domains and in physical robots.

Critical strengths

Although the paper idea is simple, they provide the supervision to build the hierarchy that is needed for planning, the empirical execution is good. Their empirical section provides a comprehensive view of the effects of the hierarchy learned in planning domains that challenge the agent in different ways.

Critical weaknesses
  • In terms of presentation, the mathematical notation is quite dense and makes the fundamental pieces of the methods hard to follow. The idea is simple, the excessive use of symbols makes it hard to parse. I would encourage the authors to simplify notation. The ideas are simple, the math should not obscure it.
  • In the empirical section, the baselines chosen include hierarchical methods that learn hierarchies without the kind of scaffolded supervision this paper proposes, which tells me that a good hierarchy is better than a bad one. But it doesn't show that this particular algorithm is better within its own class (supervised hierarchy learning).

In general, the idea of hierarchies in planning and decision-making have been around for a while and it has been repeatedly shown that when the hierarchy is structured correctly it helps long-horizon planning. The hard part is to discover the hierarchical structure during learning. Thus, providing supervision to learn the hierarchy isn't particularly novel. Though this specific solution might be useful for the people in the JEPA community.

Questions for authors

See concerns above. In order to accept this paper, I would like to see an improvement in presentation, particularly around the notation in the method. Moreover, I believe experiments should include baselines that are more closely related to this kind of supervised hierarchy learning.

3 comments
AuthorsOpenReview ↗

[Part 1 of 2 — W1 & W3: Notation Simplification & Execution-Grounded Structure]

Dear Reviewer 8NtZ,

Thank you for your constructive review and for highlighting our empirical evaluation across simulated planning domains and physical robots. Part 1 addresses notation simplification (W1) and what ExAbs contributes beyond supervising a hierarchy (W3); Part 2 presents the comparisons with supervised hierarchy learners (W2). All revisions appear in the uploaded PDF.

[W1 & Q1-a] Simplifying Mathematical Notation in Section 3

We agree that Section 3 introduced too many symbols for a straightforward idea. We rewrote Section 3 to remove clashes, drop parameter subscripts from the main text, add a compact notation box, and state each step in plain words before its equation. Specifically, we use qq for the AM completion model and MM and Φ\Phi for the macro-action and effect encoders, and simplify the high- and low-level objective notation to match the captions of Figures 1–2. Segmentation and shortlist details are provided in Appendices B–D.

Keeping all original equation numbers intact, the method now reads in four equations:

  • AM completion (Eq. 1): q(B∣z,w)=Pr⁡(T≤B∣z,w)q(B \mid z, w) = \Pr(T \le B \mid z, w).
  • Segment selection (Eq. 2): choose t0<⋯<tSt_0 < \dots < t_S maximizing ∑n=1Sstn−1,tn\sum_{n=1}^S s_{t_{n-1}, t_n} subject to dmin⁡≤tn−tn−1≤dmax⁡d_{\min} \le t_n - t_{n-1} \le d_{\max}.
  • Low-level control (Eq. 4): JL(i)=[(1−ω)Ai+ωD~i]++RiJ_L(i) = [(1 - \omega) A_i + \omega \widetilde{D}_i]_+ + R_i, with Φ\Phi distilled via MSE⁡(Δa^,ΔaT)\operatorname{MSE}(\Delta\hat{\mathbf a}, \Delta\mathbf a^T) in Eq. 5.
  • High-level planning (Eq. 6): JH(r)=CH(r;zg)−λAMlog⁡ρr+λbar[ϑ−ρr]+1−ϑJ_H(r) = C_H(r; z_g) - \lambda_{\mathrm{AM}} \log \rho_r + \lambda_{\mathrm{bar}}\frac{[\vartheta - \rho_r]_+}{1 - \vartheta}, where ρr\rho_r scores first waypoint wrw_r.

We welcome further suggestions on this proposed presentation and would be glad to refine it.

[W3] What ExAbs Contributes Beyond Supervising a Hierarchy

Supervising a hierarchy is well studied; what changes in ExAbs is where the signal comes from and what it shapes. No external subgoal labels, task stages, symbolic predicates, or rewards are provided:

  1. Planner-grounded boundaries vs. trajectory heuristics and HRL: Predictive JEPA hierarchies cut trajectories at fixed strides or predictable events, yet latent proximity predicts controller completion at only 0.55 AUROC vs. 0.79 for AM (Table S4). In HRL, HIRO and HAC relabel subgoals for fixed horizons, SoRB and LEAP score reachability from a learned policy's value, and PEAR segments demonstrations using a primitive RL policy's value (Sections 1–2). World-model planners have no actor value to query, and offline boundaries define both the macro-actions MM and prediction targets of FH\mathcal F_H. Retraining the high level on AM-selected segments alone (L2 in Table 5, with no online AM or PETAL) already beats HWM in 9 of 11 settings, raising Maze Hard from 77.5% to 82.5% before reaching 90.0% with full ExAbs.
  2. Executor-matching control (Table S20): Because PETAL alters how waypoints are pursued, AM must model that exact controller. We performed single-seed paired evaluations in each scene. Holding world models, high-level model, PETAL controller, queries, budgets, and seeds fixed, swapping only AM between Matched AM (PETAL rollouts) and Anchor-label AM (anchor-only rollouts) improves all four scenes:
SettingAnchor-label AM + PETALMatched AM + PETALGain (pp)
Counterflow50.50% (101/200)57.50% (115/200)+7.00
Reach-Wall40.63% (39/96)45.83% (44/96)+5.21
Two-room17.19% (11/64)21.88% (14/64)+4.69
Cube d=50d=5060.00% (30/50)62.00% (31/50)+2.00
Total (410 episodes)44.15% (181/410)49.76% (204/410)+5.61
AuthorsOpenReview ↗

[Part 2 of 2 — W2: Baselines Within Supervised Hierarchy Learning]

[W2 & Q1-b] Comparisons with Trajectory-Supervised Hierarchy Learners

Following your suggestion, we expanded Sections 1–2 and compared ExAbs across 12 settings against three trajectory-supervised hierarchical methods in Section 5.3 and Table S21: HIQL (Park et al., NeurIPS 2023), HSVL (Tiofack et al., ICML 2026), and HGCBC (Gupta et al., CoRL 2019). All three train high-level subgoal selectors and low-level actor policies for 500k gradient steps (batch size 256) on the same offline trajectories, frozen visual encoders, proprioception, and evaluation budgets as ExAbs.

To be explicit about data and interaction scope: HIQL, HSVL, and HGCBC train closed-loop actors purely offline from dataset transitions, whereas ExAbs trains no actor policy and plans over the frozen world model at deployment. PETAL itself is distilled from imagined planner candidates on offline states without environment execution (Appendix D.3); environment interaction enters ExAbs only through the preparation-stage attempts used to fit AM (Appendix E).

EnvironmentHIQLHSVLHGCBCExAbs (Ours)
Maze Easy85.0$^\dagger$60.0(5.0)‡^\ddagger100.0
Maze Medium62.5$^\dagger$40.0(5.0)‡^\ddagger97.5
Maze Hard42.5$^\dagger$5.0(0.0)‡^\ddagger90.0
Counterflow17.536.014.054.5
Two-room6.328.14.720.3
Reach-Wall68.825.088.545.8
Cube d=25d=2562.7 ±\pm 5.859.3 ±\pm 1.256.0 ±\pm 0.068.7 ±\pm 3.1
Cube d=50d=5046.0 ±\pm 5.366.0 ±\pm 6.054.0 ±\pm 3.559.3 ±\pm 6.4
Cube d=75d=7552.7 ±\pm 3.164.7 ±\pm 6.460.0 ±\pm 3.572.7 ±\pm 4.6
PushT d=25d=2556.0 ±\pm 4.056.0 ±\pm 4.042.0 ±\pm 14.083.3 ±\pm 3.1
PushT d=50d=5035.3 ±\pm 4.252.0 ±\pm 4.038.0 ±\pm 7.240.7 ±\pm 2.5
PushT d=75d=7530.7 ±\pm 6.446.7 ±\pm 1.226.7 ±\pm 2.324.0 ±\pm 3.5

†^\daggerFrom Table 1; other baseline entries are new runs. Cube and PushT report mean ±\pm SD over three seeds. ‡^\ddaggerHGCBC did not train reliably on Maze under its state-encoder adaptation (≤5\le 5\\%); values are shown in parentheses for completeness and excluded from comparisons.

Without training an actor, ExAbs attains the highest success in 7 of 12 settings, with its largest margins where offline dataset step counts diverge from obstacle or flow feasibility: Maze Hard (90.0% vs. 42.5%), Maze Medium (97.5% vs. 62.5%), Counterflow (54.5% vs. 36.0%), and PushT d=25d=25 (83.3% vs. 56.0%). In the remaining settings where policy-learning methods lead, these methods outperform all world-model planners in the paper, while ExAbs still clearly outperforms the HWM world-model baseline on the same backbone (e.g., Reach-Wall 34.4% →\to 45.8%; Two-room 7.8% →\to 20.3%).

AuthorsOpenReview ↗

Thank You for Your Feedback & Final Follow-up

Dear Reviewer 8NtZ,

We have responded to your comments and updated the paper accordingly. As the discussion period is nearing its end, we would be happy to answer any further questions or discuss any remaining concerns. Thank you for your time and help!

Best regards,
The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to the Area Chair and All Reviewers

We thank the Area Chair and reviewers for their careful reviews. Our revision addresses four main themes: notation and presentation, trajectory-supervised hierarchy baselines and executor matching, the mechanism behind PETAL, and task scale and hardware repeatability. In short, ExAbs lets the local controller of a latent world-model hierarchy determine both where high-level temporal units begin and end (AM, fitted to the controller's own closed-loop completions) and what each waypoint asks the controller to do (PETAL, a source-relative transition effect), without external stage labels, symbolic predicates, or actor value functions.

The table below summarizes our new experiments, ablations, and manuscript revisions:

ThemeNew Evidence & RevisionsLocation in Revised PDF
Notation & Unified FramingRemoved symbol clashes and parameter subscripts; added a compact notation box and plain-language introductions before equations; expanded comparisons with supervised hierarchy learners and neuro-symbolic methods.Sections 1–3; Figs. 1–2; Appendices A–D
Supervised Baselines & Executor MatchingEvaluated HIQL, HSVL, and HGCBC across 12 settings. Across valid comparisons, ExAbs leads in 7/12 settings without training an actor (Maze Hard 90.0% vs. 42.5%). Executor-matched AM outperforms anchor-labeled AM in all four evaluated scenes (49.8% vs. 44.1% over 410 episodes).Section 5.3; Appendix L.4; Tables S20–S21
Background Controls & Internal PETAL AblationsRescoring ground-truth physical outcomes after changing a shared background reduces base-scoring goal attainment in 3 of 5 tasks, suggesting that shared-background effects on candidate selection can arise independently of prediction error. PETAL reduces the drop in each of those three tasks. Internal ablations (T0/T1, A/R/P, FixedD, Uniform, NoSL) show that metric learning and selective spatial support can improve success, while shortlist effects are small and vary across tasks. These secondary design choices have relatively modest effects; taken together, the results suggest that PETAL's gains arise primarily from learning source-relative effect representations.Sections 1, 3.3, 5.2–5.3; Appendices D.2, K.4, L.4; Tables S16–S19
Longer-Horizon Tasks & Robot RepeatabilityAdded sequential Button–Drawer (18.3% vs. 0.0% HWM) and 8-DoF Visual AntMaze (46.0% vs. 36.0% HWM). An independent 50-trial physical evaluation of Basket Transfer reaches 30.0% (15/50), close to the original 28.0% (14/50).Sections 4.2, 4.4; Appendices H.1, H.4; Table S15

Detailed point-by-point responses are posted under each review below.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.