acceptodds
Under review as a conference paper at ICLR 2027

TrekWM: Long-Horizon Planning via Shaped Latent Geometry and Adaptive Subgoals

Abstract

Latent world models (LWMs) offer a promising framework for zero-shot visual control by planning in a latent representation space rather than raw pixels. However, latents trained exclusively for prediction do not naturally induce a metric suitable for evaluating planned trajectories: Euclidean distance between latents fails to reliably reflect true environment steps, leading to severe degradation over long task horizons. To address this, we propose TrekWM to improve long-horizon planning in LWMs via two complementary mechanisms. First, a ranking loss shapes the latent geometry so that latent Euclidean distance preserves temporal ordering when scoring candidate futures during planning. Second, we shorten the effective planning horizon using an adaptive subgoal proposer trained with a flexible temporal schedule, enabling the discovery of meaningful task landmarks rather than relying on rigid, fixed intervals. Together, these complementary improvements ensure both reliable local distance estimation and scalable global planning. Across eleven settings in five environments, TrekWM achieves the top performance in ten, outperforming competing methods by a substantial margin. The advantage is most pronounced on the longest-horizon tasks: vs. on Cube, vs. on PushT, and vs. on TwoRoom, in each case compared against the strongest baseline for that setting.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
latent world model

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_AG9gOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses zero-shot, long-horizon, goal-conditioned visual planning with JEPA-style latent world models trained on offline, reward-free data. The authors argue that latents trained only for prediction give distances that do not track progress toward distant goals. They also argue that fixed-interval subgoals blend distinct task phases. TrekWM addresses this in two ways. A pairwise ranking loss shapes the world-model embedding so that latent distance preserves temporal order. A winner-take-all subgoal proposer places subgoals at real latents within a temporal window. Across 11 settings in five simulated environments, rank-correlation analysis and a 2×2 factorial results indicate that the two components are complementary, and extensive ablations support the design choices.

Critical strengths
  1. Well-motivated problem with measured hardness. The paper addresses zero-shot, long-horizon, goal-conditioned visual planning from offline, reward-free data. It shows empirically that latents trained only for prediction stop tracking progress toward distant goals: LeWM's rank correlation with time-to-go falls to about zero beyond short horizons (Table 19). It also explains why fixed-interval subgoals blend distinct task phases.
  2. Clear positioning and a testable hypothesis. The work is positioned between JEPA world models, cost and metric shaping, and hierarchical planning. The central hypothesis is stated clearly: a plannable latent needs distances that preserve temporal order, and subgoals that keep each planning segment within the range where those distances are reliable.
  3. Design choices are empirically justified, and the results support the hypothesis. Shaping the embedding itself outperforms a separate metric head (Table 12). The flexible-schedule objective places subgoals at real bottlenecks, with doorway localisation of 99.2 vs 59.8 for the fixed schedule (Table 4). Shaping raises rank correlation in every environment (Table 19). The 2×2 factorial (Table 10) shows the components are complementary on Cube-50: success is 67.3 with both, against 32.0 with subgoals alone and 0.7 with shaping alone.
  4. Thorough ablations and appropriate metrics. The ablations cover the regulariser and ranking loss, embedding vs head, fixed vs flexible schedule, number of subgoals, clock and episode budget, pointer-check interval, HWM search budget, HIQL-style subgoals and the subgoal chain on Temporal-JEPA. Success rate is the standard metric, and the rank-correlation analysis directly tests the geometric claim.
Critical weaknesses
  1. Narrow scope and empirical-only hardness. "Long horizon" means at most 100 steps on filtered simulated tasks, and Cube episodes are truncated, so the scope is narrower than the framing suggests. The hardness argument is entirely empirical: there is no formal account of when or why distances learned for prediction stop tracking progress.

  2. Incremental novelty. Value-guided JEPA and Temporal-Distance JEPA already shape the embedding itself, so the gap reduces to combining shaping with adaptive subgoals. Each component is also close to prior work:

  • The ranking loss falls within the line of work that uses temporal supervision to shape representations, such as time-contrastive learning and VIP. The paper should spell out how ordinal ranking on the embedding differs from these methods.
  • The subgoal objective, as the authors acknowledge, adapts Multiple Choice Learning and Explorative Modeling.
  • The spectral hinge appears secondary and task-dependent (Table 11), and it performs poorly without the ranking loss on PushT.
  1. Engineering contribution without theoretical grounding. The method trains only on forward pairs, treating goal distances as asymmetric, yet the planner uses symmetric Euclidean distance. The paper does not analyse this mismatch. Results only partly support the hypothesis. Shaping alone gives little or no gain on long manipulation tasks: direct-planning success falls from 11.3 to 8.7 on PushT-100 and rises only from 0.0 to 0.7 on Cube-50 (Table 10). Shaping also degrades world-model prediction (Table 18). On TwoRoom, one-step relative error rises from 0.03 to 1.53. On Reacher, position-probe error rises from 0.63 to 3.04 cm.

  2. Insufficient statistical rigour and incomplete metrics.

  • Results use a single training seed with no confidence intervals, so several wins fall within binomial noise of about ±4 points (e.g., Maze-far 34.7 vs 33.3).
  • There is no sensitivity analysis for τ, the pair weights or the ranking gate.
  • Path efficiency, success-vs-horizon curves for all environments, compute and planning latency, and significance tests are not reported.
  1. The SOTA claim needs qualifying.
  • Weaker baseline configuration. HWM is evaluated in the authors' image-only shared setup with a reduced sampling budget, not its published configuration. The authors' own run of HWM's released code reports Diverse Maze results above TrekWM's (Section A.6). The protocols differ, so the superiority claim should be stated as holding within this shared setup.
  • Missing comparisons. The paper cites but does not compare against closely related methods (ProWorld, Value-guided JEPA, TRM, Traj-LeWM, FF-JEPA).
  • Unexplained result. Temporal-JEPA outperforms TrekWM on Reacher-50, and the paper does not explain why.
Questions for authors

Refer Weaknesses section

8 comments
AuthorsOpenReview ↗

Response to Reviewer AG9g (1/5): W1

Thank you for the detailed review. We will clarify the scope of the paper and include the analyses and comparisons below. And we will revise the paper accordingly.

W1. Horizon, task design, and planning geometry

Horizon. We use “long horizon” relative to the latent world-model baselines in these simulated tasks. Our main evaluation reaches 100 steps, compared with 25 in LeWM's protocol (Section 4 and Table 1). Table R10 extends the horizon sweep, and Table R12 reports the results on HWM's Maze distance bands.

Table R10. Success versus start–goal distance. TrekWM are means over three training seeds. “TrekWM − subgoals” uses the same models and plans directly to the goal. † indicates a significant improvement over “TrekWM − subgoals” after Holm correction.

Table R10(a). Cube

Method10 steps20 steps25 steps30 steps40 steps50 steps60 steps
TrekWM89.384.083.6†86.0†78.9†68.2†44.7†
TrekWM − subgoals96.483.865.645.124.00.90.0
TrekWM − subgoals − shaping (LeWM)77.334.026.721.310.00.00.0

Table R10(b). PushT

Method25 steps50 steps75 steps100 steps125 steps150 steps
TrekWM93.382.4†77.8†67.1†60.9†45.1†
TrekWM − subgoals90.751.334.714.78.99.6
TrekWM − subgoals − shaping (LeWM)86.736.024.011.313.310.0

Table R10(c). Reacher

Method25 steps50 steps75 steps100 steps150 steps
TrekWM90.787.690.496.999.6†
TrekWM − subgoals93.689.394.496.495.6
TrekWM − subgoals − shaping (LeWM)80.050.752.753.357.3

Table R10(d). TwoRoom

Method50 steps75 steps100 steps
TrekWM100.0†100.0†100.0†
TrekWM − subgoals97.394.480.4
TrekWM − subgoals − shaping (LeWM)24.017.316.7

Task filters and Cube truncation. Both filters use task geometry (Appendix A.1). For Cube-25, we remove windows where the block moves less than the 4 cm success radius. These account for 38% of the pool, and a random policy succeeds on 43% of the unfiltered tasks. In TwoRoom, we keep pairs in opposite rooms at least 100 units apart, so the agent has to go through the doorway. We cut Cube demonstrations after the first successful placement (about 70 frames) because they continue into a second attempt. Cube-50 still tests reaching, grasping, transporting, and placing the block (see Table R9 in our response to Reviewer pibC).

Why prediction alone leaves distances undetermined. Suppose g(f(ot),at)=f(ot+1)g(f(o_t),a_t)=f(o_{t+1}) and the latent distribution is isotropic Gaussian. A radius-dependent rotation φ\varphi preserves that distribution. Replacing the encoder with f′=φ∘ff'=\varphi\circ f and the predictor with g′(z,a)=φ(g(φ−1(z),a))g'(z,a)=\varphi(g(\varphi^{-1}(z),a)) preserves perfect prediction and the distribution targeted by SIGReg. The rotation can change Euclidean distances and reorder candidate plans. Ranking directly constrains that order. For distant goals, TrekWM retains positive correlation with time-to-go as LeWM's correlation falls (Table 19).

AuthorsOpenReview ↗

Response to Reviewer AG9g (2/5): W2 & W3

W2. Ranking, subgoal chains, and prior work

Novelty. Our contribution is how latent shaping and adaptive subgoals work together. Subgoals break the task into shorter stages, and the planner uses the shaped distances to judge progress within each stage. On Cube-50, neither component is enough on its own. Shaping alone reaches 0.7% success and subgoals alone reach 32.0%. Together they reach 67.3% (Table 10). The choice of latent space still matters when we hold the subgoal planner fixed. On Cube-50, Temporal-Distance JEPA reaches 4.7% and the two VG-JEPA variants reach 20.7% and 36.7%, compared with 68.2% for TrekWM (Table 16 and Table R6 in our response to Reviewer hXpg).

Ranking. Our loss supervises which state is closer to a shared future goal in the embedding used by the predictor, proposer, and planner. Time-contrastive learning supervises temporal proximity within a window. VIP learns a discounted value used as a reward, and VG-JEPA fits a goal-conditioned value in the latent distance. Our loss constrains the order without fitting a value scale. Temporal-Distance JEPA predicts directed time-to-go with a separate head. With the same subgoal planning framework, TrekWM performs better on harder manipulation tasks, while results are closer on Maze (Tables 12 and 16 and Table R6 in our response to Reviewer hXpg).

Subgoals. Multiple Choice Learning and Explorative Modeling inspire our closest-target matching. Our generator extends this idea to sequences by producing an entire subgoal chain in one pass. This raises two questions specific to a chain: whether to match the chain as a whole or match each subgoal separately, and how to preserve progress toward the goal without fixing the spacing between subgoals. We match each subgoal independently to the closest observed state in its own time window in the planner's latent space. The windows move forward with the subgoal index, encouraging coarse temporal order while letting the generator choose where to place each waypoint. The learned chain places more waypoints at the TwoRoom doorway than a fixed schedule. The manipulation ablations favor learned subgoals and show that the same proposer benefits from TrekWM's latent space (Tables 3, 4, 10, 16, and 17).

Spectral hinge. As stated in the submission, the hinge is a supporting regularizer that prevents collapse. Its standalone performance is weak on PushT. Ranking also improves success with SIGReg, which can be used in place of the hinge (Table 11).

W3. Symmetry, shaping alone, and prediction error

Symmetry. We agree that the symmetric Euclidean cost cannot capture directional reachability, even though the ranking pairs are sampled forward in time. Modeling this asymmetry with a quasimetric is an important direction for future work. In TrekWM, local subgoals limit how far the planner has to rely on the symmetric distance at each stage.

Shaping and subgoals. Shaping alone yields small or negative gains on long manipulation tasks, which motivated our use of adaptive subgoals (Table 10). Ranking helps the planner judge progress, and subgoals let it plan over shorter segments. Shaping helps in some flat-planning settings, and the two components together perform better on harder tasks (Table 17, Table R10, and Table R8 in our response to Reviewer pibC).

Prediction error. Prediction error can rise without changing which plan looks best. On TwoRoom, much of the error comes from a nearly constant offset. With the same flat planner, TrekWM ranks plans better and succeeds more often despite higher five-step prediction error (Table R11). On Reacher, most position-probe error is already present before rollout. Table 18 shows how prediction error varies across tasks.

Table R11. Prediction accuracy and planning-cost geometry on TwoRoom. All models use the same flat planner. FVU measures 5-step prediction error (lower is better); rank columns report Spearman correlation.

World modelFVU ↓Cost rankGeodesic rankExpert in top 10% (D = 50)Expert in top 10% (D = 100)Success (D = 50)
LeWM0.0380.7190.26241.417.024.0
LeWM, TRM training recipe0.056–0.0590.7700.51156.730.848.7
LeWM released by the TRM authors0.265–0.2860.7420.49555.725.341.1
TrekWM0.1150.9240.84598.252.697.3
AuthorsOpenReview ↗

Response to Reviewer AG9g (3/5): W4 - part 1

W4. Uncertainty, sensitivity, and efficiency

Training seeds. Table R12 reports confidence intervals over three TrekWM training seeds and paired tests against LeWM. Due to limited compute, LeWM uses one training run. The gain intervals resample TrekWM seeds and evaluation tasks with LeWM fixed. The paired tests show gains over LeWM, including on Maze.

Table R12. Three-seed results with confidence intervals and significance tests (Weakness 4). Mean ± s.d. and 95% t-interval over three training seeds. Δ compares TrekWM with LeWM: its 95% CI comes from a bootstrap that resamples training seeds and evaluation tasks, and p from a paired t-test over the three runs. * indicates that all three seeds achieve 100% success. Baselines in this table are those evaluated in the submission. For PushT-50, the CI and p of Δ use the per-task outcomes of a LeWM rerun (35.3%).

Task typeTaskTrekWM mean ± s.d.TrekWM 95% CILeWMΔ vs. LeWM [95% CI]pBest baselineΔ vs. best baseline
Same traj.Cube 2583.6 ± 13.6[49.8, 100]26.7+56.9 [+41.1, +70.7]0.019DINO-WM: 62.7+20.9
Same traj.Cube 5068.2 ± 2.7[61.5, 74.9]0.0+68.2 [+61.5, +74.7]<0.001DINO-WM: 0.7+67.5
Same traj.PushT 5082.4 ± 1.4[79.0, 85.9]36.0+46.4 [+38.2, +54.9]<0.001HWM: 41.3+41.1
Same traj.PushT 10067.1 ± 2.8[60.2, 74.0]11.3+55.8 [+47.3, +64.0]<0.001HWM: 23.3+43.8
Same traj.TwoRoom 50100.0 ± 0.0[100, 100]24.0+76.0 [+68.7, +83.0]*DINO-WM: 86.7+13.3
Same traj.TwoRoom 100100.0 ± 0.0[100, 100]16.7+83.3 [+76.0, +90.0]*T-JEPA: 54.7+45.3
Same traj.Reacher 5087.6 ± 4.2[77.0, 98.1]50.7+36.9 [+27.3, +46.4]0.004T-JEPA: 90.7−3.1
Same traj.Reacher 10096.9 ± 2.7[90.2, 100]53.3+43.6 [+35.3, +52.0]0.001T-JEPA: 93.3+3.6
Mazenear79.1 ± 1.4[75.7, 82.6]9.3+69.8 [+62.9, +76.2]<0.001T-JEPA: 52.7+26.4
Mazemid54.9 ± 5.6[41.1, 68.7]6.0+48.9 [+39.6, +58.0]0.004T-JEPA: 28.7+26.2
Mazefar30.0 ± 7.0[12.7, 47.3]3.3+26.7 [+17.8, +36.0]0.022T-JEPA: 14.7+15.3
Cross traj.Cube59.1 ± 2.1[53.8, 64.4]0.7+58.4 [+51.3, +65.3]<0.001HWM: 2.7+56.4
Cross traj.PushT40.4 ± 2.7[33.8, 47.1]8.7+31.8 [+23.1, +40.2]0.002HWM: 8.7+31.8
Cross traj.TwoRoom99.6 ± 0.8[97.6, 100]6.7+92.9 [+88.7, +96.7]<0.001T-JEPA: 44.7+54.9
Cross traj.Reacher94.7 ± 2.4[88.7, 100]50.0+44.7 [+36.7, +52.9]<0.001T-JEPA: 80.7+14.0
Reviewer_pibCOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses an important problem in latent-space planning: representations learned for prediction may not provide a distance metric that reflects progress toward a goal. TrekWM introduces temporal ranking supervision to shape the latent geometry and achieves substantial improvements on the evaluated benchmarks.

Critical strengths
  • The paper identifies a meaningful limitation of predictive world models and provides a clear motivation for learning planning-oriented representations.

  • The improvements are substantial on several benchmarks, particularly on longer-horizon tasks.

Critical weaknesses
  • This paper reminds me of Temporal Straightening [1], which seems to be very similar to TrekWM limits my confidence in the novelty of the proposed method. Discussion and direct comparisons should be included.

  • In many robot tasks, temporally distant states can be semantically similar, especially in tasks with loops or revisits. Forcing them apart in Euclidean embedding space may distort the learned representation. Evaluation on more complex, multi-stage manipulation tasks would help clarify the method’s applicability.

[1] Wang, et al. Temporal Straightening for Latent Planning. ICML 2026

Questions for authors
  • How does the proposed ranking objective compare with Temporal Straightening?
  • How does temporal ranking behave on trajectories containing revisits, waiting, or detours?
5 comments
AuthorsOpenReview ↗

Response to Reviewer pibC

Thank you for the constructive feedback, especially the connection to Temporal Straightening and the questions about revisits. We added a direct Temporal Straightening baseline, measured revisits in the training data, and analyzed detours and multi-stage control. We will revise the paper accordingly.

W1 / Q1: Comparison with Temporal Straightening.

Both methods use temporal structure to shape the latent space for planning. Temporal Straightening aligns successive latent displacements over three consecutive states. TrekWM's ranking loss compares the distances of two states to a shared future goal, which can be many steps away. Small local turns can accumulate and disrupt the ordering of distances to a distant goal, while a curved trajectory can still approach its goal at every step. The two losses could therefore complement each other by encouraging local straightness and goal-relative ordering in the same latent space.

We still compared both methods directly on the same tasks with their own planners. TrekWM outperforms Temporal Straightening in 7 of 8 same-trajectory settings (Table R8(a)). TrekWM also has higher success on all four cross-trajectory tasks, where starts and goals come from different held-out trajectories, and at all three distances on unseen Maze layouts (Table R8(b)). Without subgoals, TrekWM leads by at least 8.6 percentage points in 10 settings.

Table R8. Direct comparison with Temporal Straightening. Success rate (%) on the same- and cross-trajectory tasks and on unseen Maze layouts. TrekWM rows average three training seeds.

Table R8(a). Same trajectory

MethodCube 25Cube 50PushT 50PushT 100TwoRoom 50TwoRoom 100Reacher 50Reacher 100
Temporal Straightening41.30.742.720.764.030.790.077.3
TrekWM83.668.282.467.1100.0100.087.696.9
TrekWM − subgoals65.60.951.314.797.380.489.396.4
TrekWM − subgoals − shaping (LeWM)26.70.036.011.324.016.750.753.3

Table R8(b). Cross trajectory and unseen Maze layouts

MethodCubePushTTwoRoomReacherMaze nearMaze midMaze far
Temporal Straightening0.07.329.362.765.321.35.3
TrekWM59.140.499.694.779.154.930.0
TrekWM − subgoals0.74.062.491.379.341.819.8
TrekWM − subgoals − shaping (LeWM)0.78.76.750.09.36.03.3

W2 / Q2: Revisits, waiting, and detours.

We agree that loops, revisits, and waiting can create conflicting ranking targets. TrekWM's memoryless encoder (following lewm) cannot handle visually identical observations whose correct action depends on history; addressing this requires a history-conditioned model and is important future work.

In our training data, revisits occur in 25% of the TwoRoom trajectories and 34% of the Maze random walks. Despite this, TrekWM solves TwoRoom (Table R8). With random walks alone, it outperforms both behavior-cloning baselines on far Maze goals (the data-quality study in our response to Reviewer hXpg, W3, Table R3). In cross-trajectory TwoRoom, the agent has to go through the doorway (see Q2 and Table R7 in our response to Reviewer hXpg). TrekWM nearly solves this task and substantially outperforms both flat planning with the same models and LeWM (Table R8).

This is possible because, in these tasks, the route from a location to a fixed goal does not depend on how the agent reached that location. A revisit can make the ranking target noisy, but its transitions remain valid for learning the dynamics. At test time, TrekWM searches for new action sequences and replans from the current observation instead of repeating the loop. Together, valid transition learning and test-time replanning explain why performance remains high despite imperfect temporal supervision.

W2: Multi-stage manipulation.

Cube requires the agent to reach, grasp and lift, transport, and place the block. With subgoals, TrekWM completes 68.2% of tasks, compared with 0.9% for flat planning with the same models, and improves every prerequisite stage (Table R9).

Table R9. Stage-wise progress on Cube. Each entry is the percentage of episodes that reach the corresponding stage. “− subgoals” uses the same TrekWM models and plans directly to the goal.

MethodReachGrasp & liftTransportPlace (success)
Expert demonstration100.0100.0100.0100.0
TrekWM96.279.169.168.2
TrekWM − subgoals2.00.40.90.9
TrekWM − subgoals − shaping (LeWM)2.70.70.00.0

Thank you again for your review. With two days left in the discussion period, we wanted to check whether our response and the additional experiments addressed your concerns. If anything is still unclear or you have any new questions, please let us know so we have time to respond before the discussion closes.

Thank you for your response and clarifications. I will maintain my current score.

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

This paper studies the problem of using a learned model for planning. It suggests two concrete improvements:

  1. Fine tune the representations of the model based on a ranking loss.
  2. Train a model to propose subgoals between the agent's current and goal states. At test time, the agent uses simple traj-opt to reach waypoints that eventually lead to the goal.
Critical strengths
  • the diagnosis is strong and makes sense: there is no reason for prediction latents to be good over long horizons for planning.
  • empirical analysis and ablations are thorough.
  • Shaping representations and waypoint driven planning is a good combination for long-horizon tasks.
Critical weaknesses
  1. The ranking loss implicitly assumes that the behavior policy is a good one and sets a bad upper bound on the time-to-go between the start and goal states. There are already many methods that do this using general dynamic programming type methods by learning goal-conditioned value functions---both in the online and offline settings---so this index-based loss feels like a step back. At the very least, why not compare against the value based baselines like HIQL, QRL, or even Value Guided JEPA, which you cite.
  2. I have a similar issue with the subgoal generator, which is trained on a supervised learning loss that also implicitly works well only when the behavior policy already produces high quality data. Just like goal-conditioned value learning, this problem of proposing subgoals is extensively studied in offline and online HRL; why not compare against proper HIQL or LEAP (both of which are cited I think)?
  3. Why no comparisons to BC baselines? Given that your method implicitly assumes high quality demonstration data, you should compare against high quality BC methods like goal-conditioned BC or Diffusion Policies.
Questions for authors
  1. Questions above.
  2. Evaluations: IIUC, you sample a held-out demonstration trajectory and a start state in that trajectory. Then you sample a goal state in that same trajectory? Why? Why not sample a random start-goal pair from the environment and see if the agent can generalize to pairs that are not in the demonstrations? that is the standard of proof for methods that use demonstration data---that they can generalize outside the demonstrations as well.
6 comments
AuthorsOpenReview ↗

Response to Reviewer hXpg (1/3): Data Quality and Behavior Cloning

Thank you for the detailed and constructive feedback. In response, we added a controlled data-quality study, HIQL, QRL, Value-Guided JEPA, LEAP, GCBC, and Diffusion Policy baselines, and a new cross-trajectory evaluation. We will revise the paper accordingly.

Baseline setup. HIQL, QRL, and GCBC use the official OGBench implementations and recommended hyperparameters. Diffusion Policy uses its official network code and image-PushT hyperparameters, with the goal image as a second camera input. Due to no official implementation, we reimplemented Value-Guided JEPA. Our implementation covers both its Euclidean (VF) and quasimetric (VF quasi) variants. All methods use the same training split and evaluation tasks. We average main results over three training seeds.

W1: Data quality.

Our training data include substantial nonexpert behavior following LeWM's data protocols (Table R1 and submission Appendix Table 5). Only 40% of TwoRoom episodes and 65% of PushT episodes reach their goals. In Maze, half of the training frames come from random walks.

Table R1. Training data of each environment. Maze contains random-walk episodes; TwoRoom and PushT also contain many unsuccessful demonstrations.

EnvironmentBehavior policyEpisodesFramesGoal-reaching rate
Cubescripted pick-and-place oracle (OGBench)9,5000.64M99.7%
ReacherSAC expert, targets in sequence9,5001.90M100%
PushTdemonstrations of mixed quality17,7502.22M64.5%
TwoRoomnoisy scripted navigation9,5000.88M40.2%
Maze50% shortest-path expert, 50% random walk76,9335.39Mexpert 100%, random 5.9%

W3: Behavior cloning.

We compare TrekWM with GCBC and Diffusion Policy (Table R2). TrekWM outperforms both BC baselines on PushT and long-distance Maze goals, where the training data include substantial nonexpert behavior, while both BC baselines perform better on Cube, where the training data are near-expert.

Table R2. Comparison with goal-conditioned behavior cloning.

MethodCube 25Cube 50PushT 50PushT 100TwoRoom 50TwoRoom 100Reacher 50Reacher 100Maze nearMaze midMaze far
GCBC97.394.760.744.792.790.788.098.067.338.013.3
Diffusion Policy100.098.067.360.796.0100.074.789.390.047.326.7
TrekWM83.668.282.467.1100.0100.087.696.979.154.930.0

To further investigate how data quality affects this comparison, we train all three methods on equal-size Maze datasets with different proportions of expert demonstrations and random walks (Table R3). Both BC baselines outperform TrekWM when most of the training data come from experts. Expert demonstrations provide effective actions for BC to imitate directly, whereas TrekWM benefits from broader coverage of states and actions to learn dynamics for planning.

As the expert share drops to 50% or less, TrekWM performs better. With random walks alone, both BC baselines lose performance, while TrekWM improves over its expert-only result and still performs best. These results show that TrekWM can learn useful dynamics even from failed trajectories, reducing its reliance on expert demonstrations.

Table R3. Success rates (%) by expert data share for far goals in Maze. Each setting uses half of the full Maze training dataset. TrekWM's world model and subgoal proposer are both trained on all of these frames.

Method100% expert75% expert50% expert25% expert0% expert
GCBC42.724.710.715.30.0
Diffusion Policy75.335.319.319.33.3
TrekWM4.78.737.330.722.0
AuthorsOpenReview ↗

Response to Reviewer hXpg (2/3): Value-Based Baselines and Subgoal Planning

W1: Value-based baselines.

We compare TrekWM with HIQL, QRL, and Value-Guided JEPA (Table R4). TrekWM outperforms all three on Cube, PushT, and unseen Maze layouts.

Table R4. Comparison with offline RL and Value-Guided JEPA.

MethodCube 25Cube 50PushT 50PushT 100TwoRoom 50TwoRoom 100Reacher 50Reacher 100Maze nearMaze midMaze far
HIQL49.362.040.017.3100.0100.0100.0100.04.70.70.0
QRL69.32.02.70.784.096.795.398.038.712.76.7
LEAP (offline)2.70.01.30.0100.0100.059.348.06.72.00.7
Value-Guided JEPA (VF)61.32.022.07.3100.098.090.094.068.736.728.7
Value-Guided JEPA (quasi)61.31.344.010.780.755.384.780.733.316.77.3
TrekWM83.668.282.467.1100.0100.087.696.979.154.930.0

These results show that a simple ranking loss can support effective planning even with non-expert data. The Maze random-walk data help explain why. Moving between two states took about 3 times as long as taking the shortest route. Even so, the later state was closer to the goal in 79.7% of pairs.

We then tested whether the learned distances capture progress in held-out demonstrations. For each pair of states five frames apart, we checked whether the method ranks the later state closer to the same future goal. TrekWM does so more often than the value-based baselines for distant goals on Cube and PushT (Table R5). Temporal ranking directly trains the ordering that our planner uses to compare predicted outcomes without requiring an accurate numerical estimate of the cost to reach the goal.

Table R5. Progress toward distant goals on held-out demonstrations. Frames left are counted from the first state to the goal. Scores show how often (%) the state 5 frames later ranks closer to that goal.

MethodCube (30–60 frames left)PushT (75–150 frames left)
HIQL94.776.7
QRL54.580.6
VG-JEPA (VF)92.487.6
VG-JEPA (VF quasi)76.054.8
TrekWM (ranking)97.493.0

W2: Subgoal learning and hierarchical RL.

Expert trajectories can make subgoal learning easier by providing effective intermediate states, but in principle they are not required because the generator learns to propose states observed within trajectory segments. Consistent with this, TrekWM still leads both BC baselines on far Maze goals when trained on random-walk data alone (Table R3).

We compare with HIQL and train LEAP offline using its official code (Table R4). LEAP struggles on Cube, PushT, and Maze, where its VAE often fails to reconstruct the object or agent being controlled.

We also implement a hierarchical version of VG-JEPA using TrekWM's subgoal planner so that we can compare the two world models under the same planning framework. With this shared planner, TrekWM performs better (Table R6). This result suggests that how we shape the latent space matters even when both models use the same subgoal planner.

Table R6. World-model comparison with the same subgoal planner. All rows use TrekWM's subgoal planning framework.

World modelCube 25Cube 50PushT 50PushT 100TwoRoom 50TwoRoom 100Reacher 50Reacher 100Maze nearMaze midMaze far
VG-JEPA (VF) + subgoal73.320.718.712.0100.0100.080.088.072.749.332.7
VG-JEPA (VF quasi) + subgoal88.736.766.043.396.798.784.090.740.018.010.0
TrekWM83.668.282.467.1100.0100.087.696.979.154.930.0
AuthorsOpenReview ↗

Response to Reviewer hXpg (3/3): Generalization Beyond Demonstrations

Q2: Generalization beyond the demonstrations.

The original same-trajectory protocol is the standard evaluation used by latent world-model planners: LeWM, DINO-WM, Temporal Straightening, Temporal-Distance JEPA and so on. Also, the submission already tests randomly sampled start–goal pairs on 20 unseen Maze layouts. TrekWM leads on long-distance goals (Table R7).

To test generalization beyond a single demonstration, we also paired starts and goals from different held-out trajectories (Table R7). Compared with the longer same-trajectory setting, cross-trajectory success rates stay similar on TwoRoom and Reacher and drop on Cube and PushT. Even with these drops, TrekWM leads the world models shown on all tasks.

Table R7. Generalization beyond individual demonstrations. Success rates (%) are shown as same / cross. Same uses the longest evaluated setting for each task. DINO-WM is not evaluated on Maze due to its planning cost (Appendix A.3).

MethodCubePushTTwoRoomReacherMaze nearMaze midMaze far
DINO-WM0.7 / 0.013.3 / 5.343.3 / 33.374.0 / 64.0−−−
PLDM0.0 / 0.74.7 / 0.753.3 / 38.063.3 / 50.730.017.36.7
LeWM0.0 / 0.711.3 / 8.716.7 / 6.753.3 / 50.09.36.03.3
Temporal-JEPA0.0 / 0.012.7 / 4.754.7 / 44.793.3 / 80.752.728.714.7
HWM0.0 / 2.723.3 / 8.723.3 / 21.352.0 / 41.320.04.76.7
TrekWM (ours)68.2 / 59.167.1 / 40.4100.0 / 99.696.9 / 94.779.154.930.0

Summary.

TrekWM learns useful dynamics from nonexpert trajectories and uses temporal order to guide planning even when those trajectories are inefficient. With subgoal planning, TrekWM achieves the strongest overall performance among the latent world models we compare, leading on all four cross-trajectory tasks and across all three goal distances on unseen Maze layouts.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank all reviewers for their careful reviews and helpful suggestions. Their main concerns fall into three groups: comparisons with policy methods and the need for expert data, generalization and harder tasks, and comparisons with other latent world models. Below, we summarize the key evidence and revisions. Changes are marked in blue in the revised PDF.

1. Does TrekWM require expert demonstrations? How does it compare with behavior cloning and value-based RL? (hXpg, AG9g)

No. TrekWM can learn a useful planning model from nonexpert data. Our new experiments show when this is an advantage.

  • Behavior cloning and data quality. TrekWM outperforms GCBC and Diffusion Policy when the training data are of mixed quality. To see how data quality affects this comparison, we varied the share of expert data in equal-size Maze datasets. BC does better with mostly expert data, while TrekWM does better when the expert share drops to 50% or less. With random walks alone, TrekWM reaches 22.0% success on far goals versus 3.3% for Diffusion Policy, showing that TrekWM can learn to plan from failed trajectories. (Reply to hXpg, W3.)
  • Value-based RL. We compared TrekWM with HIQL, QRL, and offline LEAP. TrekWM ranks first in nine of eleven settings across five environments. We updated Table 1 and narrowed the performance claims to reflect these differences. (Reply to hXpg, W1 and W2.)

2. Does TrekWM generalize well and perform well on harder tasks? (hXpg, pibC)

Yes. Cross-trajectory tests show that TrekWM can connect starts and goals from different held-out demonstrations and performs best among the tested world models in all environments. (Reply to hXpg, Q2.) We also added three further tests.

  • Cube generalization. To test generalization on Cube, we changed the start state so the gripper was closed. The gripper starts open in every training episode. Both BC baselines dropped to 0% success, while TrekWM still reached 31%. (Table R16.)
  • Farthest Maze goals. On the farthest goals in unseen layouts, TrekWM reaches 24.7%, compared with 12.7% for Diffusion Policy, 9.3% for GCBC, and 2.0% for LeWM.
  • Higher-dimensional manipulation. On LIBERO with 7-D actions, TrekWM reaches 29.6% when starting 100 steps before completion, while LeWM reaches 0.0%. The environments we previously use are the shared suite of prior latent world-model planners, chosen for direct comparison; scaling beyond it is future work, with LIBERO as a first step. (Second-round reply to hXpg.)

3. Does TrekWM improve over other latent world models, and what explains the difference? (all reviewers)

Yes. We added comparisons with both Value-Guided JEPA variants, Temporal Straightening, TRM, and Traj-LeWM. Among all latent world models in Table 1, TrekWM performs best in ten of eleven settings across five environments.

With the same subgoal planner, TrekWM reaches 68.2% on Cube-50 versus 36.7% for the best Value-Guided JEPA variant, suggesting that how we shape the latent space also matters (Table R6).

Other revisions.

We added three-seed results, confidence intervals, paired tests, sensitivity checks, path lengths, and compute costs. We expanded the analysis of shaping and subgoals, clarified the training data and HWM comparison, and narrowed the performance claims. The appendix gives further results, implementation details, and limitations.

Summary.

The reviewers highlighted the following strengths:

  • Problem and motivation. All three reviewers recognized the problem we study: distances learned for prediction can stop tracking progress toward distant goals (hXpg, pibC, AG9g). The analysis also explains how fixed-interval subgoals can mix different stages of a task (AG9g).
  • Method and results. The reviewers saw value in combining latent shaping with waypoint planning (hXpg, AG9g). They also noted the clear presentation and substantial gains on several benchmarks, especially at longer horizons (pibC).
  • Evidence. hXpg and AG9g noted the thorough experiments and ablations. AG9g also highlighted the clear positioning and testable hypothesis, backed by success rates, rank correlation, the metric-head comparison, and doorway localization.

Our contribution is to learn both where to aim and how to judge progress in the same world-model embedding. The generator learns adaptive subgoals from observed states in flexible time windows. Temporal ranking shapes the distance that the planner uses to choose actions and decide when to move to the next subgoal. On Cube-50, success reaches 67.3% with both components, compared with 32.0% with subgoals alone and 0.7% with shaping alone in the same ablation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.