acceptodds
Under review as a conference paper at ICLR 2027

Learning to Plan from Random Exploration

Abstract

Random exploration reveals how an environment can be traversed before a goal is specified. Can this experience support long-range planning without policy-improvement training? Our random-walk analysis explains what temporal relations contain: short horizons reveal geodesic geometry in the diffusion limit, while longer horizons reveal connectivity between regions before mixing removes these distinctions. We learn these relations with a conditional energy-based model that estimates temporal log-density ratios through horizon-conditioned embeddings. The model is trained on observation pairs by noise-contrastive estimation, without action or reward labels. The planner queries these learned relations at different horizons as it moves toward the goal. At test time, a separate local dynamics model predicts candidate action outcomes, and the temporal model evaluates their progress toward the goal by selecting or aggregating estimated improvements across horizons. The agent executes one action and replans with both models fixed. Experiments demonstrate long-range maze planning from random exploration using states and images. Learned score fields, embedding probes, and planned routes exhibit properties of a multiscale cognitive map. We further demonstrate egocentric navigation from random exploration and manipulation planning from suboptimal data.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
random exploration, representation learning, goal-directed planning, energy-based models, self-supervised learning, cognitive maps

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_UovyOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes to learn an energy-based model to quantify the distance between latent states in JEPA-style world models. They show that when this model is used for planning, it enables achieving goals even when the dynamics model is trained with data collected with random actions. There is a theoretical analysis of the proposed approach and the model is evaluated on several environments, with a focus on navigation tasks.

Critical strengths
  • Theoretical analysis: while I found the paper hard to follow, there is a theoretical motivation that helps to understand what the authors try to achieve a bit better.
  • Diverse environments.
  • The main insight that an EBM that quantifies the distance to goal can help planning when the dynamics are learned from random data is interesting (though not entirely convincing).
  • Seemingly better results on data collected with random actions.
  • Extensive appendix.
Critical weaknesses
  • Writing: the paper, starting from the abstract, was not easy to follow. The abstract is just the culmination of sentences that do not provide motivation or informative descriptions of the proposed method. The paper does not have a very coherent storyline. A lot of terms are used without explanation, for example (L42): “capture regional connectivity before mixing erases these distinctions.” What is regional connectivity? What does mixing refer to? L88: “positive prediction horizons”. What are positive prediction horizons?. This trend continues throughout the entire paper. My impression is that I’m reading a very concise output of an LLM. In addition, the sections are probably organized in the wrong order, as the method section mentions a lot of terms that are found later, e.g., “the free goal head” (L114). Lines 160-161: sentence is cut half-way through, no connector to the following one.
  • All implementation details are deferred to the appendix for some reason, so the main text jumps from theory to experiment without any practical connection.
  • Notations: hθ,gθh_{\theta}, g_{\theta} use the same parameters θ\theta. Does it mean they are the same network with different outputs or are they different networks? Unclear.
  • “Exploration”: I am not sure I understand what this paper has to do with exploration. There is no interaction with the environment, the proposed approach learns purely from offline data that might be collected with random actions, and learns an EBM to give an estimation for the distance from the goal for planning. To my understanding, the authors try to train the baselines with random data (that means that for LeWM for example, the dynamics model is trained random transitions) and the results show that they are not usable for planning. So the insight here is that if the dynamics model does not cover a wide range of behaviors, it is not usable for training, which is not surprising. I’m not convinced that the additional EBM overcomes that coverage issue.
  • The related work section in the main text is very lacking and barely covers existing exploration methods.
Questions for authors
  • My main concern is that if I understood the paper, then the novelty here is very limited, and the pompous theoretical analysis around it does not really help the case. I admit that perhaps I did not understand the paper very well as I found the writing very hard to follow. I believe the paper would benefit from a significant writing revision.
4 comments

Consider an agent randomly walking through a maze without knowing where its future goal will be. Although its actions are random, its movements are constrained by the environment, so the resulting trajectories reveal how different locations are connected.

We do not learn the random policy or directly fit a goal-distance function. Instead, we learn multi-horizon temporal relations from observation pairs (xt,xt+τ)(x_t,x_{t+\tau}), estimating how likely one observation is to follow another after τ\tau steps relative to its marginal frequency:

Gθ(x,y,τ)≈log⁡pdata(y∣x,τ)p0(y).G_\theta(x,y,\tau)\approx \log\frac{p_{\mathrm{data}}(y\mid x,\tau)}{p_0(y)}.

Our conditional EBM learns these relations through NCE, without action labels or shortest-path supervision. Our theoretical analysis further explains how these temporal statistics relate to geodesic geometry and regional connectivity.

Different horizons reveal different structures. Proposition 1 shows that short-horizon temporal scores recover geodesic geometry in the diffusion limit, while Proposition 2 characterizes the spectral structure of longer-horizon scores, explaining how slowly decaying modes preserve information about connectivity between regions. At sufficiently long horizons, however, the random walk approaches its stationary distribution, making the temporal relations increasingly independent of the starting location. This theoretical analysis explains why different horizons capture complementary spatial information and motivates multi-horizon planning. Fig. 1 illustrates these properties:

  • Fig. 1A: At τ=16\tau=16, the temporal score concentrates near the goal; at τ=512\tau=512, it extends through connecting passages.
  • Fig. 1B: At τ=1\tau=1, embeddings preserve local geodesic geometry; at τ=4096\tau=4096, they reflect connectivity between distant regions.
  • Fig. 1C: Planned routes exhibit adaptive horizon selection, generally using shorter horizons as the agent approaches the goal.

W2: How does this enable planning?

Once a goal is specified, the agent needs to decide which action moves it closer to that goal. We separate this into two components:

  1. Local dynamics: Predict the one-step outcomes of candidate actions.
  2. Temporal model: Evaluate the predicted outcomes using their learned temporal relations to the goal across multiple horizons.

The agent selects an action, executes it, and replans. Unlike dynamics-based world-model planners, our method does not require multistep dynamics rollouts. Long-range guidance is directly provided by the learned temporal relations, while dynamics prediction is needed only for local action selection.

Importantly, our contribution is not simply adding an EBM as a goal-distance metric. We learn horizon-conditioned temporal statistics directly from experience, rather than fitting a supervised distance function. These learned relations provide long-range planning guidance without explicitly predicting intermediate states.

W3: Why random exploration?

We agree that our use of "exploration" may have been unclear. We do not propose or learn an exploration policy. Rather, we use exploration in the behavioral sense: collecting experience without a task-specific goal or reward, with random walks being the simplest example. The resulting trajectories are used for offline learning. Our question is whether such goal-free experience can later support goal-directed planning. We have clarified this distinction in Section 2.1 and expanded the discussion of exploration-policy methods in Section 5.

We also agree that sufficient data coverage is necessary, and do not claim that our method overcomes missing coverage. However, coverage alone does not guarantee effective long-range planning. To examine this, we train all methods on the same 4M-step random-walk dataset, which visits every free cell:

MethodLargeGiant
Best offline RL baseline0.98 (HIQL)0.27 (QRL)
GP / PAP (ours)1.00 / 1.000.80 / 0.92

Despite using the same exploratory data, the baselines struggle on Giant, while our method achieves substantially higher success. This suggests that how temporal structure is learned from the data matters beyond state coverage alone.

W4: Clarity, notation, and implementation

We agree that several concepts and implementation details were insufficiently explained, and have made targeted revisions to address the specific issues raised.

Clarity and organization. The Introduction now explains "regional connectivity" as the existence of feasible paths between regions, and "mixing" as the loss of dependence on the starting location as a random walk approaches its stationary distribution. Section 2.1 now defines the horizon τ\tau as the number of steps between observations. We have also corrected the incomplete sentence at L160–161.

Notation. Section 2.2 now explicitly states that hθh_\theta and gθg_\theta share a backbone with source and target heads. The heads are tied for navigation and separate for manipulation.

Implementation. The learning objective and planning procedure are described in Section 2, with detailed implementation settings provided in Appendix D. We have clarified the connection between these components in the main text.

Related work. Section 5 now distinguishes our approach from JEPA-style predictive learning and exploration-policy methods, with extended discussions in Appendix A.

We sincerely appreciate the reviewer's feedback, which has helped us improve the paper. We hope these clarifications and revisions address the concerns, and we would welcome any further questions.

Reviewer_UovyOpenReview ↗

I don't believe this paper is ready for publication

I thank the authors for their effort during the rebuttal period. I went over the revised paper and the other reviews and responses as well. My impression has, unfortunately, not changed - I find the paper barely readable and my impression is that it is mostly LLM-generated (apologies if not, but this is my impression). I also don't understand how the first part of the response is related to my review. The seemingly better results do not justify publication as writing is a major part of how we communicate findings and propose ideas. From my understanding, the novelty here is very limited and the usage of "exploration" is quite misleading. I'll keep my score.

We thank the reviewer for revisiting the revised manuscript. We have taken the writing concerns seriously and made concrete revisions to the terminology, notation, organization, and implementation details.

Our previous response directly addressed the reviewer's concerns about exploration and novelty. We would like to restate two central points to make this connection more explicit.

1. Random exploration is the source of experience, not the policy we aim to learn.

Our central question is how goal-free exploratory experience can be converted into reusable knowledge for goal-directed planning. We do not propose an exploration policy or claim to overcome insufficient data coverage. The fact that learning is performed offline does not change the exploratory nature of the experience from which the data were collected.

More importantly, random trajectories are not merely coverage data. Although actions are random, the resulting trajectories are constrained by the environment. Their multi-horizon temporal statistics encode information about the underlying geometry and connectivity. Our matched-data experiments show that learning and using this temporal structure matters even when all methods train with the same data.

2. Our method learns multi-horizon temporal relations, rather than directly fitting goal distance.

We learn horizon-conditioned temporal relations through NCE, without action labels or shortest-path supervision. Our theoretical analysis characterizes how different horizons capture complementary geometric and connectivity information, motivating their use in planning.

We further develop planning algorithms that leverage these learned multi-horizon temporal relations. They combine long-range guidance from the temporal model with one-step dynamics for local action selection, without requiring multistep dynamics rollouts.

These distinctions directly address the reviewer's original concerns about the characterization and novelty of our method. Our contribution lies in learning, theoretically characterizing, and developing planning algorithms that utilize multi-horizon temporal structure from random exploratory experience.

Finally, regarding the impression that the manuscript is mostly LLM-generated, we respectfully disagree. We have addressed the specific writing concerns through concrete revisions. We believe the manuscript should be evaluated based on its clarity, technical formulation, theoretical analysis, and empirical evidence.

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

“Learning to Plan from Random Exploration”, as the name suggests, learns long-term planning behavior from trajectories collected using random walks. An energy-based model is trained with contrastive loss to estimate the log-density ratio between the conditional probability of a future state given a current state and a horizon, and the marginal probability of the same future state in the dataset. This temporal model is trained to estimate how strongly a future state is temporally associated with a current state at a given horizon. Theoretical analysis shows the temporal relation in short horizon (geodesic distance) and long horizon (connectivity of the environment). A lightweight one step dynamics model is also learned to predict the outcome of each candidate action, whose progress towards a given goal state is then evaluated by the temporal model. Experiments are primarily conducted on a PointMaze environment with two sizes (Large and Giant) and compared to other offline goal conditioned RL and world model baselines with the proposed method achieving strong performance in this benchmark.

Critical strengths
  1. The paper grounds the method well in theory, explaining that short horizons encode geodesic distance while longer horizons encode connectivity between regions. Fig. 1 shows this clearly with scores (Fig. 1 A) and embeddings (Fig. 1 B) shifting from local geometry to regional connectivity as horizon increases. Fig. 1 C also shows the planner moving to shorter horizons when nearer to the goal.
  2. The results are strong. On random exploration data for the Large and Giant PointMaze, the proposed method reaches 0.85 to 0.97 success, while most baselines achieve little to no success. The authors also note an interesting observation that their proposed method performs well on a manipulation (Cube) benchmark with suboptimal data, convincingly beating LeWM with cleaner demonstration data. It is also in interesting that on the same manipulation benchmark, using noisy data alone (0.908 success rate) outperforms expert-only data (0.876).
Critical weaknesses
  1. The evaluations are not fully matched with the temporal model using 4M lattice transitions while reproduced baselines use only 1M random-action transitions. Furthermore, the dynamics model needs a further 1M or 10M action-labelled transitions. While Table 2 in the results section (4.2) shows the proposed method staying competitive at 1M lattice transitions used, a lattice-based dataset is likely to yield higher coverage and data diversity, which favours the proposed method.
  2. Results are primarily on PointMaze, and the Habitat experiments feel like an afterthought where results are only reported in prose against a random policy, with no explanation of why the position-only planner turns in place. Manipulation experiments on the Cube benchmark did not include Offline RL methods like GCIQL, which have decently strong performance on this benchmark. Furthermore, manipulation experiments do not use random exploration data which is the central thesis of the paper, and instead uses expert demonstrations mixed with noisy-oracle data.
Questions for authors
  1. Why is the proposed method trained on lattice transitions while the other baselines in Table 1 use uniformly sampled actions, and are those actions discrete or continuous? Table 6 in the appendix shows that with matched 1M lattice data, HIQL reaches 1.00 success on Large for both official tasks and random pairs, largely closing the gap. Why is the lattice used only for the proposed method, and how do the methods compare on Giant PointMaze under matched lattice data with 4M transitions.
  2. Table 2 in the results section (4.2) shows the success rate of the proposed method scaling with the size of the dataset used, with 1M transitions resulting in 0.49 to 0.50 success rate and 4M transitions resulting in 0.76 to 0.79 success rate. Given that the Giant PointMaze is still a relatively simple 2D maze, how does the required data scale while environment size or state dimension? Is the approach practical for larger environments where random exploration covers the space much more slowly?
5 comments

We sincerely thank the reviewer for the thoughtful and insightful questions. We have taken these concerns seriously and conducted additional experiments and analyses to address them carefully. We respond to each point below.

W1/Q1: Data fairness

We agree that the submitted Table 1 was not matched. In the revision every method in a comparison is trained on the same data, as we will elaborate below.

(1) Giant under matched 4M lattice data. The revised Table 1 trains the offline RL baselines on exactly the 4M lattice walk our models use; they observe its positions and take its moves as actions. Three training seeds.

Official tasks, SR (SPL)LargeGiant
GCIQL0.97 (0.96)0.00 (0.00)
QRL0.96 (0.88)0.27 (0.24)
HIQL0.98 (0.97)0.23 (0.22)
HILP0.58 (0.56)0.02 (0.02)
GP (ours)1.00 (0.95)0.80 (0.63)
PAP (ours)1.00 (0.96)0.92 (0.78)

On Giant random pairs, GP/PAP reach 0.88/0.90 against 0.53 for QRL and HIQL.

(2) Does the lattice favour our method? The lattice walk helps every method. On Large, GCIQL, QRL and HIQL reach 0.96 to 0.98, close to our planners, so Large no longer separates the methods once the data cover it well; the revised text says this. The gap is on Giant, where routes are long: on 100 held-out Giant pairs, baseline success is 0.85 to 1.00 for shortest paths under 20 units and 0.07 to 0.31 for paths of 60 to 80 units.

(3) Random-action data. The actions are continuous: a uniformly sampled 2-D action at every simulator step. The new Table 2 trains every method on the same random-action data at 1M and 4M transitions (and on the lattice walk at both sizes). At 4M:

Official tasks, SRLargeGiant
GCIQL0.370.00
QRL0.650.12
HIQL0.000.00
HILP0.400.00
GP (ours)0.930.47
PAP (ours)1.000.61

Our method depends on the data as well. Random actions diffuse slowly, so a horizon covers a much shorter distance than on the lattice. We state this in Section 4.2.

(4) Latent Dynamics Model. The local dynamics model is separate from the temporal model and servers only to predict the one-step outcome of each candidate action, so that the planner can choose among them. Long-range guidance comes only from the temporal model, which is trained on the same 4M transitions as the baselines. The lattice walk has no continuous actions, so for the lattice columns the dynamics is trained on separate random-action transitions (1M on Large, 10M on Giant, Section 4.1). The larger dataset for Giant (10M) is used solely to improve the one-step prediction accuracy, rather than to provide additional long-horizon guidance.

To check that this extra data is not where our results come from, we replace the learned dynamics with the simulator, which gives oracle one-step outcomes (Appendix Table 6). The results agree within seed variation:

Official tasks, SRLarge, GPLarge, PAPGiant, GPGiant, PAP
Learned dynamics0.960.750.760.88
Oracle dynamics (simulator)1.000.700.730.87

On random-action data the comparison is fully matched: the dynamics is trained on the same 4M transitions as every other model, with nothing extra, and our planners still lead (0.93/1.00 on Large and 0.47/0.61 on Giant, against at most 0.65 and 0.12 for the baselines).

W2: Habitat and manipulation

Habitat. The Habitat results are reported in Appendix Table 10, and Section 4.2 now explains the failure of the position-only planner. This baseline observes the position only, with neither the heading nor the image. A turn therefore leaves its observation unchanged, so the planner cannot tell a turn toward the goal from a turn away from it, and it keeps turning. This failure reflects the missing heading information rather than a limitation of the temporal representation. The first-person image carries the heading, which is why image plus position reaches the goal.

Manipulation. Unlike navigation, which evaluates planning toward arbitrary goals, manipulation focuses on a fixed task with a more constrained goal distribution. We therefore make a practical choice to use task-relevant trajectories rather than relying exclusively on random exploration. To investigate this choice, we collect trajectories from uniformly random robot actions, without goal or expert guidance, and mix them with expert data:

Expert data ratio0%25%50%75%100%
Episodes where the cube moves8.2%31.1%54.2%77.1%100%
PAP0.7000.8770.8900.8850.876

Pure random exploration rarely produces meaningful object interactions (8.2% of episodes), although it still supports some planning. Adding expert data substantially improves performance. Our manipulation experiments therefore address a complementary question: whether the same temporal model can plan from suboptimal trajectories (expert mixed with noisy-oracle data, Table 4). This distinction is explicitly stated in the abstract and Section 4.1, and our framework does not require random exploration as a training assumption.

Baseline comparison. We follow LeWM's data and evaluation protocol and compare against the reported baselines. We also reproduce GCIQL under this protocol, obtaining a success rate of 0.678, compared with approximately 0.64 reported in the LeWM paper, and update Table 4 accordingly. We do not claim an advantage over offline RL methods on manipulation.

Q2: Scalability

We agree that naive random walks become increasingly data-inefficient in larger environments. Table 3 distinguishes state coverage from the coverage needed to learn long-horizon relations. A 0.1M-step walk already visits every free cell in Giant, yet GP improves from 0.01 to 0.76 when the trajectory length increases from 0.1M to 4M steps. That is, visiting every state is not sufficient: reliably estimating long-horizon conditional relations requires repeated trajectories connecting distant regions. This limitation is expected for diffusive exploration, where the typical displacement from the starting point grows only with the square root of the number of steps.

Our central contribution is a framework for learning long-range planning representations from goal- and reward-free trajectories. Importantly, the learning objective is not restricted to uniformly random walks and can accommodate trajectories from other behavior policies (e.g., suboptimal manipulation trajectories). The framework also extends beyond 2D state observations, as demonstrated by our image-based navigation and manipulation experiments. We leave systematic scaling to higher-dimensional states and larger environments to future work, potentially with more efficient exploration strategies.

We sincerely appreciate the reviewer's thoughtful questions, which have helped strengthen our work. We welcome further questions and would be grateful if the reviewer would consider a higher rating.

Sorry for the late response, appreciate the work

For the latent dynamics model, it needs an extra 10M data. It is also critical for your method to function as a planner. Are these statements correct? Comparing it with a simulator does not change the fact that your method requires an additional 10M data points in an offline RL setting, could you justify this additional 10M steps.

Scalability is a huge issue, because while there are guarantees for random walk, they do not scale well to the AntMaze for Large and Giant. In such high dimensional systems, random walk to produce a gait for traversal costs an astronomical amount of time. Could the authors comment on whether the given analysis and planner can extend to such settings?

Reviewer_u5TqOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies the control problem whether the agent is capable of learning to plan based on randomly collected trajectories. Existing methods often rely on policy learning or multistep dynamics rollouts. The paper learns multi horizon temporal relations through contrastive estimation and combines them with local dynamics predictions for goal directed planning. Experiments evaluate the approach on maze navigation, egocentric navigation, and manipulation tasks.

Critical strengths
  1. clean design. The method separates longrange temporal guidance from local dynamic predictions provides a nice approach that capable of plan.
  2. theoretically sound. The theory shows why learning multi horizon is necessary that potentially capture complementary information.
  3. The method was tested across multiple environments, with clean ablations.
Critical weaknesses
  1. It appears to me that the proposed method share highly similar high level idea with V-JEPA, or other JEPA based planning methods, that they are all learning representations from trajectories to support goal directed planning. From my point of view, the method use scoring instead of latent prediction as used by general JEPA method. I want to hear the author's input on how their proposed method distinct with these works, especially the difference between scoring and prediction.

  2. Inconsistency in data. I noticed that the paper use 1M for large and 10M for giant for baselines (section 4.1) whereas they also say their method used 4M transitions, which seems like a not fair comparison.

Questions for authors

Please try to address the weakness I listed.

3 comments

We thank the reviewer for the insightful and positive assessment. We address the two concerns below.

W1: Distinction from JEPA

Thank you for raising this important question. The key distinction is how long-range information is represented and used for planning.

(1) Learning: JEPA-based world models learn action-conditioned predictive latent dynamics, whereas our method directly learns multi-horizon temporal relations from observation pairs through NCE, without requiring action labels for representation learning. Specifically, predictive dynamics models learn local action-conditioned state transitions, while our scoring model directly captures long-range temporal relations at different horizons, without explicitly predicting intermediate states.

(2) Planning: This leads to a different planning paradigm. Dynamics-based planners typically evaluate distant goals by repeatedly applying learned dynamics to roll out candidate action sequences. In contrast, our model directly scores the temporal relation between each candidate next state and the goal at any learned horizon. This separates long-range guidance from local action prediction: the temporal model provides long-range guidance, while a one-step dynamics model predicts candidate action outcomes. No multistep dynamics rollout is required.

(3) Theory: Our analysis in Appendix B further characterizes how these temporal relations capture local geodesic geometry and regional connectivity at longer horizons, providing a theoretical interpretation of the learned representation.

We discuss this distinction in the world-models paragraph of Related Work (Section 5).

W2: Data fairness

Thank you for the question. The submitted Table 1 compared our models on 4M lattice transitions with baselines on 1M random-action transitions. In the revision, every method in a comparison is trained on the same 4M data.

Table 1 (revised): all methods on the same 4M lattice walk. The offline RL baselines observe the positions of the walk and take its moves as actions. Three training seeds.

Official tasks, SR (SPL)LargeGiant
GCIQL0.97 (0.96)0.00 (0.00)
QRL0.96 (0.88)0.27 (0.24)
HIQL0.98 (0.97)0.23 (0.22)
HILP0.58 (0.56)0.02 (0.02)
GP (ours)1.00 (0.95)0.80 (0.63)
PAP (ours)1.00 (0.96)0.92 (0.78)

With matched data the baselines solve Large as well. But on more complex Giant Maze our model outperforms other baseline significantly.

Table 2 (new): random-action and lattice data at 1M and 4M transitions, again with identical data for all methods. For example, on 4M random-action transitions:

Official tasks, SRLargeGiant
GCIQL0.370.00
QRL0.650.12
HIQL0.000.00
HILP0.400.00
GP (ours)0.930.47
PAP (ours)1.000.61

Latent Dynamics Model. The local dynamics model is separate from the temporal model and has one job: it predicts the one-step outcome of each candidate action, so that the planner can choose among them. Long-range guidance comes only from the temporal model, which is trained on the same 4M transitions as the baselines. The lattice walk has no continuous actions, so for the lattice columns the dynamics is trained on separate random-action transitions (1M on Large, 10M on Giant, Section 4.1). We use 10M on Giant only to make the one-step prediction accurate.

To verify that our results are not driven by the additional training data, we replace the learned dynamics with the simulator, which gives oracle one-step outcomes (Appendix Table 6). The results are consistent within seed variation:

Official tasks, SRLarge, GPLarge, PAPGiant, GPGiant, PAP
Learned dynamics0.960.750.760.88
Oracle dynamics (simulator)1.000.700.730.87

On random-action data the comparison is fully matched: the dynamics is trained on the same 4M transitions as every other model, with nothing extra, and our planners still lead (0.93/1.00 on Large and 0.47/0.61 on Giant, against at most 0.65 and 0.12 for the baselines).

We hope these clarifications and additional experiments address the reviewer's concerns. We would be happy to address any further questions and would be grateful if the reviewer would consider a higher rating.

I appreciate the response and additional evidence. I am keeping my score.

Thank you again for your positive assessment and for the questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.