acceptodds
Under review as a conference paper at ICLR 2027

Beyond Single Goal Images: Learning Goal Metrics for LeWorldModel Planning

14 views

Abstract

A single goal image shows both task-relevant information and incidental visual details, although many visually different outcomes may solve the same task. We extend LeWorldModel (LeWM) to plan toward sets of successful outcomes by separating predictive representations from goal-specific cost evaluation. We freeze the encoder and action-conditioned latent dynamics and learn a goal projector, used as either a linear map or a nonlinear network. The distance between a projected latent vector and the mean of projected goal examples defines the planning cost. Training combines compactness of successful examples with distribution regularization, temporal ordering along goal-reaching trajectories, and preferences for demonstrated action sequences over corrupted alternatives. The resulting cost integrates naturally into the existing latent model-predictive-control procedure without modifying the world model. Across three world-model seeds, the nonlinear projector achieves mean success rates of %, %, and % on Push-T, Two-Rooms, and OGBench-Cube, respectively. It surpasses both single-goal-image LeWorldModel planning and latent 1-nearest-neighbor goal-set planning, improving over the latter by , , and percentage points under the same evaluation protocol. These results support learning task-specific planning costs while preserving a shared predictive representation. Source code will be made available.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
world models, goal-conditioned planning, goal sets, model predictive control

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_z1s5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies planning when multiple observations correspond to the same successful outcome. The authors keep the LeWorldModel encoder and dynamics frozen and train a projector to map successful observations toward a common center. Distance to this center is then used as the planning cost. The training includes goal compactness and a regularizer on ordinary observations. The full nonlinear model also uses temporal and action preferences from successful trajectories. The authors evaluate linear and nonlinear projectors on Push-T, Two-Room, and OGBench-Cube, comparing them with costs based on a single goal image or a set of goal images under the same planner. The nonlinear model improves success on all three environments, with the largest gains on Push-T and Two-Room.

Critical strengths
  1. The paper points out a problem with matching a single goal image. The planner may try to match details such as the agent position even when they do not determine success. Learning a separate cost allows the world model to keep these details for predicting dynamics.

  2. The MLP reaches 90.8% success on Push-T and 100.0% on Two-Room, compared with 75.7% and 73.3% for latent 1-NN, across three world-model seeds. The image-only linear projector reaches 81.0% on Push-T. This is a useful result since it does not use the trajectory supervision of the full model.

Critical weaknesses
  1. The full model uses successful trajectories and action preferences, while the main goal-set baselines only use goal images. Removing the two trajectory losses drops Push-T success from 90.8% to 47.5%. To me, this comparison does not show what is gained by mapping successful states to a common center, beyond learning a cost from the additional supervision. Learning visual objectives from successful examples has already been studied in [1], which the paper discusses. I would like to see a scalar cost network trained with the same data and comparable capacity. The current ablations also keep goal compactness throughout, so they do not answer this question.
  2. Each environment uses only one task definition. This does not really test the motivation of using the same predictive model for different tasks, since only one goal head is trained per environment. The gains are also mostly on Push-T and Two-Room. On OGBench-Cube, the MLP reaches 52.3%, compared with 47.7% for the best reported k-NN variant. I would find experiments with several different goals on the same frozen world model more informative.
  3. The trajectory losses assume that later states are better and demonstrated action sequences are preferable to sampled or perturbed alternatives. A different action sequence could also succeed, particularly when several outcomes are valid goals. The authors acknowledge this, but do not check how often these preferences agree with simulator outcomes. The preference loss matters to the result, since removing it lowers Push-T success from 90.8% to 80.8%.
  4. Figure 10 changes the initial agent position and measures planning success. This changes the control problem as well, so it does not directly show that the cost ignores irrelevant changes in agent position. Figure 9 also aggregates costs over states. I would like to see the costs for different agent positions at the same block pose, including held-out successful poses and nearby failures.

[1] Few-Shot Goal Inference for Visuomotor Learning and Planning, CoRL 2018.

Questions for authors
  1. How does the method compare with a scalar cost network using the same supervision and comparable capacity? What happens if goal compactness is removed while the trajectory losses are kept?
  2. Can different goal heads be trained on the same frozen world model for several task definitions?
  3. How often do the temporal and action preferences agree with simulator outcomes?
  4. How much does the cost vary with agent position when the block pose is fixed? Does it still distinguish successful poses from nearby failures?
Reviewer_zAunOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper extends leworldmodel to tasks where goals are defined by sets of images. It learns a task-specific projector by using MSE between projected states and goal, SigReg, temporal ordering and action preferences. During planning, the cost is the distance to the projected goal center instead of the single image used in prior world models.

Critical strengths
  • Simple architecture and idea
  • Ablate some design choices on the projector and cost function
Critical weaknesses
  • Single task setup: Unlike other world models that are evaluated on multiple goal/task definitions, there's only one fixed task per environment. This does not demonstrate the advantage of using a world model across different tasks. The authors should at least provide an offline BC policy trained on the same goal reaching trajectories as a baseline.
  • Strong assumptions on data: The training trajectories are not guaranteed to be the optimal solutions. The ordering in successful trajectories (which only evaluates in the end) doesnt necessarily indicate task progress, and the actions selected by the training data collection policy are possible to be worse than the perturbated ones. The robustness of the method to imperfect trajectories remains untested. Moreover, if the training trajectories can satisfy the assumptions and there's only one task, training a BC policy seems to be a more efficient approach.
  • The current setup trains one projector per task, and requires 1000 demonstration images. The advantage of the proposed approach over learning a scalar value function is unclear.
  • It has three hyperparameters in the training objectives, which make tuning quite costly.
  • The current method and eval are limited to one single world model (lewm), and it's unclear if it can be applied to other world models.
Questions for authors

see weakness

  • Could the authors provide (i) policy and (ii) value function baseline?
  • Could the authors show the robustness of the method to imperfect trajectories?
  • Could the authors apply the proposed methods to more world models? For example,

TD-MPC2: Scalable, Robust World Models for Continuous Control, ICLR 2024

DINO-WM: World Models on Pre-trained Visual Features enable Zero-shot Planning, ICML 2025

A Case for Planning with Latent Dynamics Models, NeurIPS 2025

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

The paper extends LeWorldModel to plan toward multiple valid outcomes rather than one specific goal image. It freezes the encoder and dynamics and trains a task-specific projector, scoring predicted futures by their distance to the mean projected embedding of successful examples. Training combines goal compactness, anti-collapse regularization, temporal ordering, and action preferences. The planner itself is unchanged. Experiments on diff tasks across three world-model seeds show higher success rates than baselines.

Critical strengths

The idea of separating prediction from goal evaluation is good. A task can have many successful outcomes, so matching one exact goal image can be unnecessarily restrictive. The learned goal cost addresses this directly and improves over both single-image and nearest-neighbor baselines.

Critical weaknesses
  • The biggest issue is task flexibility. Vanilla LeWorldModel can switch goal images without retraining, while this method trains a separate projector for each task (becaue it needs to use the success status). Since the experiments only test one task per environment, it is unclear how much practical task reuse the method offers
  • The loss has several task-specific weights. The paper ablates removing loss terms, but does not show how sensitive the results are to their values. This makes it hard to judge how much tuning is needed.
Questions for authors
  • For a new task in the same environment, do you need to retrain the projector?
  • How were the loss weights selected, and do nearby values give similar results?
  • Why use the mean of projected successful examples as the target? have you checked whether unsuccessful states can also receive low cost near that center?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.