acceptodds
Under review as a conference paper at ICLR 2027

Traj-LeWM: Path-Aware World-Model Planning via Latent Trajectory Cost

Abstract

LeWM is a lightweight visual world model that learns latent dynamics end-to-end from pixels and ranks candidate action sequences by the distance between their predicted endpoints and the goal. However, LeWM has two limitations. First, during training, it learns local next-step transitions without evaluating complete trajectories relative to the task goal. Second, during planning, it ranks candidates solely by predicted endpoint distance. Because model predictions may differ from actual execution outcomes, the candidate whose predicted endpoint is closest to the goal may not perform best when executed in the environment. The evolution of the complete predicted trajectory can therefore provide complementary information beyond endpoint distance. To address these limitations, we propose Traj-LeWM, which retains LeWM’s local-dynamics objective and endpoint score while introducing a goal-conditioned latent trajectory cost (LTC) that aggregates trajectory-level information as a complementary signal. During training, LTC-based trajectory-preference supervision complements next-step prediction in shaping the shared representation. During planning, LTC is combined with endpoint distance to incorporate intermediate-path information into candidate ranking. With joint endpoint-plus-LTC scoring, Traj-LeWM outperforms LeWM on Push-T, OGBench-Cube, Reacher, and Two-Room by 3, 14, 7, and 7 percentage points, respectively. Controlled experiments and ablations further verify the complementary roles of trajectory-level representation shaping and path-aware candidate ranking.

Primary area:
applications to robotics, autonomy, planning
Keywords:
World Models, Latent-Space Planning, Trajectory-Level Representation Learning, Goal-Conditioned Planning, Model-Based Control, Joint-Embedding Predictive Architectures

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_L37pOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work addresses two correlated issues in the training and inference of LeWM: during training, the model learns only target-agnostic and local state transitions; during planning, the plan was selected by the distance between the predicted endpoint latent to the target latent, while ignore the process in between. The paper proposes to train a lightweight scorer jointly with the WM, the latent trajectory cost (LTC), which evaluates the complete latent trajectory and can therefore distinguish candidates with identical endpoints. LTC shares the encoder with the WM, and is trained by a pairwise preference loss. The negative training trajectories are from handcrafted construction and real failures during the training, with expert trajectories as positives. The training of LTC also shapes the encoder. At planning time, LTC is added to the endpoint distance to rank candidates. The experiments show superior performance over LeWM on Push-T, OGBench-Cube, Reacher, and Two-Room.

Critical strengths

This work is well-motivated, addressing a concrete limitation of reward-free latent MPC: only the endpoint is accounted in planning. The proposed scoring module is simple, and appears applicable to other JEPA-style WMs. Ablations are carefully designed: Table 4 separates the effect of trajectory supervision on the representation from the effect of LTC at planning time. Table 3 adds a matched (in capacity and supervision) control with paired confidence intervals. In addition, mining negative trajectories from the model's own failures and using them to train the scorer is a sensible design.

Critical weaknesses

The main argument of this work, that scoring the complete trajectory rather than only the endpoint drives the planning gains, is not yet sufficiently supported. Table 3 provides evidence that a LTC scorer performs better than endpoint scorer. However, Table 6 shows that the compared endpoint scorer is as weak as a simple euclidean cost (comparing rows 1,4 in Table 6, almost same success rate, the learned one is even a bit worse). Table 3 therefore shows that LTC outperforms an endpoint scorer that provides essentially no benefit. Recent work shows that learned endpoint costs can be highly effective (e.g., TRM [1]). A stronger endoint scorer baseline is needed to support the central claim.

[1] Li, Liangyu, et al. "Beyond Euclidean Proximity: Repairing Latent World Models with Horizon-Matched Trajectory Reachability Metrics." arXiv preprint arXiv:2605.22164, 2026.

Questions for authors

Could you strengthen the endpoint scorer baseline, for example with a TRM-style cost or the temporal cost of TD-JEPA [2]? I see these are just arXiv works, but they would provide a stronger test of whether path information adds value beyond a well-trained endpoint cost.

[2] Bai, Jiaxin, et al. "Temporal-distance jepa: Plan-aware representation learning for latent world model predictive control." arXiv preprint arXiv:2607.25337, 2026.

3 comments
AuthorsOpenReview ↗

TRM-style and TD-JEPA-style endpoint costs with and without LTC

Thank you for raising the question of endpoint baseline strength. In Table 3 of the original submission, we compared endpoint and trajectory scorers matched in capacity and trained on the same preference pairs. As you noted, Table 6 shows that adding our original preference-trained endpoint head to the Euclidean endpoint cost had little effect on closed-loop success. Following your suggestion, we evaluated TRM-style and TD-JEPA-style endpoint costs to test whether adding LTC improves closed-loop success when the learned endpoint cost is held fixed.

The TRM-style head learns an endpoint cost by predicting how many time steps separate two observations from the same episode. For the TD-JEPA-style cost, we used the authors’ released implementation of the directed distance head and temporal loss. Forward time gaps within a trajectory serve as training targets, and goals from other trajectories provide negative examples. In our experiments, both endpoint heads score a candidate using only its predicted terminal latent state and the goal latent state. TRM · TD-JEPA

For each new endpoint cost, we held the Traj-LeWM checkpoint and trained endpoint head fixed. On each task, we evaluated the endpoint cost alone and with LTC added on the same 128 start–goal queries. Both conditions used the same CEM random seeds and planning budget. We calibrated the LTC weight on the validation set and kept it fixed during testing. The table reports the number of successful closed-loop trials.

Endpoint costLTC addedPush-T /128Cube /128Reacher /128Two-Room /128Total /512 (success rate)
Euclidean distanceNo1028394113392 (76.56%)
Euclidean distanceYes10689109113417 (81.45%)
TRM-style endpointNo808688128382 (74.61%)
TRM-style endpointYes10286101128417 (81.45%)
TD-JEPA-style endpointNo1048996128417 (81.45%)
TD-JEPA-style endpointYes10792108128435 (84.96%)

Used alone, the TD-JEPA-style endpoint cost yielded more successes than Euclidean distance on all four tasks. Its total of 417/512 matched that of Euclidean distance with LTC. Keeping this learned endpoint cost fixed and adding LTC increased the total to 435/512, a net gain of 18 successes (+3.52 percentage points). Adding LTC to the TRM-style endpoint cost increased the total from 382/512 to 417/512, a gain of 6.84 percentage points.

On Reacher, path scoring improved closed-loop success with all three endpoint costs. Adding LTC yielded 15, 13, and 12 additional successes with the Euclidean, TRM-style, and TD-JEPA-style endpoint costs, respectively. For the TD-JEPA-style cost, the remaining six additional successes came from Push-T and Cube, with three on each task. On Two-Room, the learned endpoint costs alone accounted for the full gain. Euclidean distance solved 113/128 queries. Both learned endpoint costs solved all 128 queries on their own and maintained that result when LTC was added.

Thank you for the additional experiments, which address my main concern. As the gains over the strongest endpoint cost are mostly pronounced on certain task (Reacher), I will keep my score of weak acceptance.

Thank you for taking the time to review our response and additional experiments, and for acknowledging that these experiments address your main concern.

Your suggestion to include stronger endpoint cost baselines was particularly valuable. It helped us more thoroughly validate the additional contribution of trajectory scoring beyond strong endpoint costs and better understand its effectiveness across different tasks.

Thank you again for your constructive feedback throughout the review process and for the time and effort you have devoted to evaluating our work.

Reviewer_zD3UOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Traj-LeWM, an extension of LeWM that incorporates a learned latent trajectory cost (LTC) into both representation learning and model-based planning. The motivation is that LeWM relies on next-step prediction during training and endpoint distance during planning, neither of which explicitly captures the quality of a complete trajectory relative to a task goal.

The proposed LTC aggregates latent states, temporal changes, goal-relative representations, and normalized temporal positions to produce a scalar trajectory cost. The authors introduce pairwise preference supervision using synthetic negative trajectories, including goal-mismatched and endpoint-preserving perturbed trajectories, as well as failures collected through closed-loop execution. During planning, LTC is combined with endpoint distance to rank candidate action sequences generated by the CEM.

Experiments on Push-T, OGBench-Cube, Reacher, and Two-Room show improvements over LeWM, with reported success-rate gains of 3, 14, 7, and 7 percentage points, respectively. The paper also includes controlled scorer comparisons, ablation studies, trajectory-ranking diagnostics, and a preliminary evaluation on a physical Franka FR3 robot.

The paper addresses a relevant problem and presents a coherent approach with encouraging empirical results. However, I have concerns about the methodological novelty, whether the proposed diagnostics establish genuine path-quality reasoning, and whether the experimental results sufficiently disentangle the contributions of trajectory-level representation learning and planning-time scoring.

Critical strengths
  1. Clear motivation and problem formulation. The paper identifies two limitations of endpoint-based world-model planning: the lack of explicit trajectory-level, goal-conditioned supervision and the inability of endpoint distance to distinguish trajectories with different intermediate evolutions. These limitations are clearly formulated and motivate the proposed method.
  2. A simple and practical approach. LTC can be integrated into an existing world model and CEM planner without replacing the original dynamics predictor or candidate generator. The approach is relatively lightweight, adding fewer than one million parameters to the backbone.
  3. Useful controlled experiments. The paper goes beyond comparing the complete method against LeWM. In particular, the frozen-backbone experiments compare LTC with a parameter-matched endpoint scorer under shared preference supervision and failure examples. LTC improves success rates on all four tasks, with paired confidence intervals excluding zero on Cube, Reacher, and Two-Room.
Critical weaknesses
  1. The methodological novelty of LTC is not sufficiently established. The proposed trajectory cost is defined as an average of per-step MLP outputs conditioned on the current latent state, its temporal difference, its displacement relative to the goal, and its normalized temporal position. The resulting cost is trained using a standard pairwise logistic preference loss. Although this design is reasonable, the paper does not sufficiently establish what fundamentally distinguishes LTC from existing trajectory-cost learning, preference-based learning, and goal-conditioned value-learning approaches. And The per-step aggregation itself is relatively straightforward, and the preference objective is not novel.
  2. The experiments do not conclusively establish that LTC learns useful intermediate-path information. The endpoint-preserving perturbation experiment fixes the initial and final latent states and adds Gaussian noise to intermediate states. LTC assigns a higher cost to the perturbed trajectories in all 200 tested pairs per task on Push-T, Cube, and Reacher, and in 96% of pairs on Two-Room. However, the perturbed latent trajectories need not correspond to physically realizable or dynamically valid action sequences. Consequently, LTC may learn to detect latent-space anomalies or unnatural state transitions rather than distinguish between dynamically valid trajectories with different task outcomes. For example, two trajectories may reach the same goal, while one takes a collision-prone route and the other follows a longer but safer route. A genuinely path-aware cost should prefer the latter when appropriate. The current perturbation experiments do not directly demonstrate this capability.
Questions for authors
  1. Methodological novelty: What is the principal technical contribution of LTC beyond applying existing trajectory-preference learning to LeWM? Why is the proposed per-step MLP aggregation necessary compared with simpler trajectory-cost alternatives?
  2. Evidence of path reasoning: Can the authors demonstrate that LTC correctly ranks dynamically valid trajectories with matched initial and terminal conditions but different intermediate paths? How can the proposed diagnostics rule out latent-space anomaly detection or richer endpoint reasoning as alternative explanations?
3 comments

1. What motivates LTC, and why use a per-step MLP?

LeWM learns latent dynamics through action-conditioned next-step prediction. Planning requires selecting an action sequence that reaches a goal. Two candidates can have predicted endpoints close to the goal while following different paths. We therefore retain LeWM’s next-step prediction objective and add goal-conditioned preference supervision over complete latent trajectories.

LTC is inspired by the form of the classical action functional S[q]=∫L(q,q˙,t) dtS[q]=\int L(q,\dot q,t)\,dt, which aggregates local contributions along a path. At each step, the scorer takes the current latent state ztz_t, the state change zt+1−ztz_{t+1}-z_t, the goal-relative displacement zg−ztz_g-z_t, and the normalized time t/(T−1)t/(T-1). It averages the resulting costs over the trajectory. The analogy concerns the form of the functional. Trajectory preferences determine what receives a low cost.

We use a standard pairwise preference objective. Our contribution is to use the same goal-conditioned trajectory cost for representation learning and action selection. Whereas a goal-conditioned value function can assess the current state relative to a goal, LTC scores the path predicted for a particular candidate action sequence. During training, the preference loss and dynamics prediction loss jointly shape the shared encoder. During planning, LTC supplements endpoint distance when ranking candidates.

Why a per-step MLP? An MLP allows the cost of a given state change to depend on where it occurs, the direction to the goal, and the stage of the trajectory. We tested this choice by training a linear path head and an MLP path head on the same frozen world model, with identical path inputs, preference examples, training budgets, and temporal aggregation.

The first table reports how often each head ranked a successful candidate ahead of a failed one when their executed endpoints were close. We scored candidates before execution, averaged over matched pairs within each query, and then weighted eligible queries equally. Ties received half credit.

TaskEligible queries / 128Linear head (%)MLP head (%)Gain (pp)
Push-T2365.278.3+13.1
Cube2449.3049.47+0.17
Reacher2127.1436.90+9.76
Two-Room7449.0258.46+9.44

Does the choice of head affect task completion? We integrated each head into CEM and evaluated closed-loop planning on all 128 queries per task under the same planning budget. The table reports successful executions, with C0C_0 denoting endpoint distance.

TaskC0C_0 onlyC0C_0 + linear headC0C_0 + MLP head
Push-T102/12898/128113/128
Cube83/12880/12886/128
Reacher94/128101/128111/128
Two-Room113/128115/128120/128
Total392/512 (76.56%)394/512 (76.95%)430/512 (83.98%)

With the same path information available to both heads, replacing the linear head with the MLP yielded 15, 6, 10, and 5 additional successes on Push-T, Cube, Reacher, and Two-Room, respectively. Across all tasks, the MLP completed 36 more queries than the linear head (+7.03 percentage points) and 38 more than endpoint distance alone (+7.42 percentage points). Both the ranking comparison and the closed-loop results support our choice of a per-step MLP.

2. Can LTC identify the successful candidate before execution when two trajectories start from the same state and reach nearby endpoints?

The reviewer points out that perturbing intermediate latent states may produce trajectories that cannot be executed. We therefore tested LTC on action sequences executed in the simulator. For each start–goal query, we generated 128 candidate action sequences and scored their predicted trajectories before execution. We then executed each candidate from the same initial state and matched candidates by their executed physical endpoints, retaining pairs in which one candidate succeeded and the other failed.

Holding the world model and candidate pairs fixed, we compared endpoint distance alone (C0C_0) with endpoint distance plus LTC. A pair received one point if the successful candidate had the lower predicted cost and half a point for a tie. We averaged first over matched pairs within each query and then equally over queries with at least one matched pair.

TaskEligible queries / 128Success–failure pairsEndpoint distance (%)Endpoint distance + LTC (%)
Push-T239764.277.1
Cube2415949.249.7
Reacher214228.135.1
Two-Room7456449.956.9

On Push-T, adding LTC improved ranking accuracy from 64.2% to 77.1% across 97 matched pairs. Two-Room had the broadest coverage. Across 564 pairs whose executed endpoints were at most 2 pixels apart, accuracy rose from 49.9% to 56.9%. Tightening the tolerance to 1 pixel retained 33 eligible queries and 108 matched pairs, and LTC still improved accuracy by 6.16 percentage points. These results directly address the concern about artificial perturbations. On executable candidates with closely matched endpoints, the predicted path helped distinguish successful from failed executions before either outcome was observed.

Could richer endpoint reasoning explain the improvement? We also compared LTC with a learned endpoint head that had access to the start, the full predicted endpoint vector, and the goal. On endpoint-matched Two-Room candidates, LTC’s ranking accuracy was 3.91 percentage points higher. Table 3 of the original submission provides a complementary closed-loop comparison. Both heads used the same frozen world model, matched parameter counts, identical preference supervision, shared failure examples, and the same CEM budget.

TaskC0C_0 + learned endpoint headC0C_0 + LTCLTC gain (pp)Paired 95% CI (pp)
Push-T561/640 (87.66%)573/640 (89.53%)+1.88[−1.25, +5.00]
Cube425/640 (66.41%)465/640 (72.66%)+6.25[+1.56, +10.94]
Reacher382/640 (59.69%)422/640 (65.94%)+6.25[+1.56, +10.94]
Two-Room602/640 (94.06%)628/640 (98.13%)+4.06[+0.63, +7.50]

With capacity and supervision matched, LTC yielded 40 additional successes on each of Cube and Reacher and 26 on Two-Room. The paired intervals excluded zero on all three tasks. Thus, the endpoint-matched diagnostic shows that LTC can improve ranking among executable candidates, and Table 3 shows that access to the intermediate path can improve closed-loop success even when the endpoint scorer is learned.

How much of the gain comes from training versus planning-time scoring? Table 4 first compares LeWM and jointly trained Traj-LeWM while holding the endpoint scoring rule fixed. It then holds the trained Traj-LeWM fixed and enables LTC during planning. Entries are closed-loop success rates (%).

Training methodPlanning scorePush-TCubeReacherTwo-RoomMean
LeWMEndpoint distance9674868785.8
Traj-LeWMEndpoint distance9880849489.0
Same Traj-LeWMEndpoint distance + LTC9988939493.5

With endpoint scoring fixed, mean success rose from 85.8% to 89.0% after joint training. Enabling LTC on the same trained model raised it further to 93.5%. The contributions varied by task. Cube improved at both stages (74% → 80% → 88%). On Two-Room, the gain was already present after joint training (87% → 94% → 94%). On Reacher, the gain emerged when LTC was enabled during planning (86% → 84% → 93%).

We appreciated your suggestion to test the scorer on plans we could actually execute. We think the broader lesson for world-model planning is that the path a model predicts can help the planner choose between actions whose endpoints look equally promising. We’d welcome your thoughts on this interpretation when you’ve had a chance to look over our response.

Reviewer_WetNOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper add a goal-conditioned latent trajectory cost (LTC) to the LeWM. It average a per-step MPL along a latent trajectory. It is trained with Bradley–Terry pairwise loss against two kinds of negative: synthetic negatives and closed-loop execution failure. During planning time, LTC is combined with endpoint distance to incorporate intermediate-path information in to candidate ranking.

Critical strengths
  1. comparison is carefully controlled. Table 3,6,7 fixed the setting and report paired crossed-bootstrap intervals. LTC outperformance the baseline, which is a good evidence.
  2. The limitation of the method is transparent. the author openly mention that they used extra interaction and the perturbation test may reflect OOD detection rather than true path reasoning.
Critical weaknesses
  1. Unfair comparison: Traj-LeWM gets to interact 200 times per epoch and use ground-truth success label. However, every single baseline they compare with is strictly offline with no environment interaction. To prove the method actual works, author need to show an offline-only version of Traj-LeWM to give other baseline with same interaction budget.
  2. Data Leakage? It is not clear to me whether 200 mining queries per epoch overlap with the evaluation. If there are any overlap, the LTC could fit failures on test queries.
Questions for authors
  1. Can you conduct extra experiment to address Critical Weaknesses #1?
  2. DA-LeWM (https://arxiv.org/pdf/2608.18746) reports gains over LeWM with offline auxiliary heads only, a short discussion would be helpful.
4 comments

We thank the reviewer for pointing out the difference in training interaction budgets. After each training epoch, the full Traj-LeWM executes 200 start–goal queries and uses their outcomes to construct trajectory preferences. The original LeWM receives no such feedback. The main comparison therefore measures the performance of the complete systems. To determine whether Traj-LeWM can also improve planning under the same interaction budget as LeWM, we evaluated a fully offline variant.

1. Offline comparison under equal interaction budgets

Traj-LeWM-syn retains LeWM’s next-step prediction objective and jointly trains the shared encoder and the latent trajectory cost (LTC) using synthetic trajectory preferences constructed solely from offline demonstrations. It makes no environment queries during training and uses no environment-provided success labels. At planning time, it combines predicted endpoint distance with LTC to rank candidate action sequences.

The first two rows below are from Table 4 of the original submission. The third row reports our additional evaluation of Traj-LeWM-syn with endpoint distance and LTC enabled, using 100 trials per task. All values are success rates (%).

Training methodPlanning scorePush-TCubeReacherTwo-RoomMean
LeWM (original Table 4)Endpoint distance9674868785.8
Traj-LeWM-syn (original Table 4)Endpoint distance9679809487.3
Traj-LeWM-syn (additional evaluation)Endpoint distance + LTC9681879489.5

With zero training-time environment interactions for both methods, Traj-LeWM-syn achieved a mean success rate of 89.5%, compared with 85.8% for LeWM.

Appendix Table 6 examines the contribution of the offline-trained scorer while holding the world model fixed. In this experiment, we froze LeWM’s encoder and dynamics predictor and trained the scoring heads using only synthetic preferences. The three offline conditions are reproduced below. Each entry gives the number of successful executions over five training seeds evaluated on the same 128 test queries (640 executions per task and condition).

Planning scoreScoring-head supervisionPush-TCubeReacherTwo-Room
Endpoint distance C0C_0No scoring head561/640434/640388/640598/640
C0C_0+ learned endpoint headSynthetic preferences only561/640428/640385/640598/640
C0C_0+ LTCSynthetic preferences only565/640434/640408/640622/640

With the same backbone and synthetic supervision, LTC yielded 4, 6, 23, and 24 more successful executions than the learned endpoint head on Push-T, Cube, Reacher, and Two-Room, respectively. Relative to C0C_0 alone, the differences were 4, 0, 20, and 24. The largest observed gains on the frozen backbone were on Reacher and Two-Room. The additional Table 4 result evaluates the complete offline method, including joint training, while Table 6 isolates the effect of the scorer on a fixed world model.

2. Overlap between failure-mining and evaluation queries

The main experiments follow LeWM’s evaluation-query sampling protocol. To check whether failure mining reused evaluation queries, we compared the mining and final evaluation records within each task and dataset version. We identified each query by its episode ID, start timestep, and goal timestep. We found no exact start–goal query matches in the records we checked.

3. Relation to DA-LeWM

We thank the reviewer for recommending DA-LeWM. We will discuss this work in the Related Work section of the revised manuscript. DA-LeWM shows that even when task states can be decoded accurately from latent representations, latent endpoint distance may still rank candidate plans poorly. The authors examine this by comparing rankings based on predicted endpoint cost with rankings based on task cost after execution. They use offline demonstrations to train an inverse dynamics head and a goal-conditioned action prediction head. Both heads are removed at test time, and CEM selects plans using the Euclidean distance between the predicted endpoint and the goal. DA-LeWM

DA-LeWM highlights an important point for interpreting our results: auxiliary supervision during training can improve closed-loop performance even when the planning score stays fixed. Traj-LeWM-syn uses goal-conditioned trajectory preferences, which can shape the shared representation and also train a cost that scores predicted trajectories during planning. This distinction motivates the two comparisons in Table 4. Holding the endpoint scoring rule fixed examines how trajectory-preference training affects planning, a question shared with DA-LeWM. Holding a trained model fixed and enabling LTC examines the additional contribution of intermediate-path information to candidate ranking.

Thank you for suggesting a fully offline comparison. For researchers working with fixed datasets, our results show that the same demonstrations used to train a world model can also teach its planner which paths to prefer. We’d be interested to hear what you make of the offline comparison once you’ve read our response.

Thank you for the additional experiment, discussion and clarifications. These address all my concerns, and I will raise my rating to weak accept.

We sincerely thank you for taking the time to carefully review our additional experiments and clarifications. We are glad that our response has addressed your concerns.

We also greatly appreciate your constructive feedback throughout the review process, which has helped us strengthen the experimental evaluation and improve the clarity of our work.

Thank you again for your thoughtful assessment and support.

Comments on the paper

1 comment

Dear reviewers,

Thank you for your reviews, and thank you to the authors for their responses. Please read the authors’ rebuttal and discuss whether it addresses your main concerns. Please update your review and rating if warranted, and briefly explain your current assessment, including any concerns that remain unresolved. If your assessment is unchanged, please explain why.

Thank you for your continued engagement.

Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.