acceptodds
Under review as a conference paper at ICLR 2027

Squeeze to Plan: Learning Dense Visual Representations for Efficient Planning via Bisimulation Information Bottleneck

Abstract

Visual foundation models offer robust perceptual priors for robotic manipulation, but it is computationally expensive to plan over their high-dimensional outputs. This inefficiency arises from the fact that for any given task, foundation models tend to capture irrelevant information. For example, a robot performing a pick-and-place need not reason about the visual texture of the table on which it operates. This leads us to ask – how can we develop planners which both benefit from these robust visual priors, and remain fast enough for real-time planning? Towards this goal, we introduce Squeeze-to-Plan (S2P), a framework for compressing the outputs of frozen foundation visual models into compact, task-specific state representations suitable for real-time planning. S2P achieves this using a novel Bisimulation Information Bottleneck objective, to encourage the compressed representations to discard task-irrelevant visual information while retaining task-relevant reward and dynamics information. S2P solves two real-time visual manipulation tasks where prior methods for planning in pretrained visual feature spaces generally fail, while achieving an orders-of-magnitude reduction in time needed to generate a plan. Additionally, we show that S2P's compression does not sacrifice the task-relevant perceptual priors despite its large compression factor, achieving similar robustness to prior methods against test-time perturbations to lighting and object-colour. Our results show that our method can enable efficient planning with pretrained visual features, while preserving their robustness to certain visual changes.

Primary area:
applications to robotics, autonomy, planning
Keywords:
robot manipulation, model-based reinforcement learning, bisimulation, information bottleneck

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_fuu8OpenReview ↗
Rating
3(+1)
Confidence
1(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Squeeze-to-Plan (S2P) addresses the cost of planning over high-dimensional frozen visual features by pooling DINOv3 features and proprioception into a 512-dimensional state vector. It trains this representation using reward prediction, latent dynamics prediction, and latent norm regularization, and combines MLP world models with TD-MPC2 value learning and MPPI planning. Experiments on three modified ManiSkill3 tasks report one to two orders of magnitude faster planning than the evaluated DINO-based baselines, improved success under simulated planning delays, and robustness to three appearance perturbations.

Critical strengths
  1. The planning-efficiency improvement is substantial and directly measured: Global pooling permits compact MLP rollouts instead of repeatedly predicting a patch-token state; the latency reduction is reported; observation encoding plus the complete planner call on the same L40S GPU are measured.

  2. The complete system demonstrates useful closed-loop control under a time constraint. With a 10-second simulated execution budget, Table 2 reports 62.5% LiftPeg success versus 0% for all baselines, and 37.5% PickCube success versus 7.5% for the strongest baselines, averaged over four visual settings.

Critical weaknesses
  1. S2P trains online, while the baselines train offline on a fixed 1.5-million-transition dataset (§5.3; Appendices A.3–A.4). S2P also explicitly includes TD-MPC2 value learning and policy-proposed trajectories, whereas the baseline adaptation documents a reward head without establishing comparable terminal-value or proposal mechanisms (§4.2; Table 6). With evaluation horizons of only 1–5 steps, these differences could materially affect control, and matched-data/planner controls and non-RT success rates are needed to separate learned control quality from latency penalties. This is a substantial attribution gap.

  2. Appendix A.1 provides a valid stochastic variational lower bound, but fixed nonzero Gaussian covariance still requires expectations over encoder noise. Moreover, for Eq. (6) as written, rescaling the latent by a factor c ∈ (0,1), with compensating rescaling of the dynamics and reward predictors, preserves its information and reward predictions while reducing both dynamics MSE and the latent norm penalty by c². Thus, a smaller norm need not imply removal of irrelevant information, and the paper does not connect this surrogate to the bisimulation aggregation condition behind Eq. (3).

  3. The contribution of the proposed objective beyond established compact world-model learning is insufficient. For example, DeepMDP (Gelada et al., 2019) already learns representations through reward and next-latent prediction, while TD-MPC2 (Hansen et al., 2024) combines latent prediction, reward/value learning, and online planning.

  4. The reported uncertainty is insufficient for the smaller performance and ablation claims. Figure 4 and Tables 2–3 use one training seed, so their SEMs describe evaluation variability rather than variability in learning. Results also select the best among up to 54 planner configurations using only 10 evaluation episodes per baseline configuration and 30 for S2P, without a separately described confirmation evaluation.

Questions for authors
  1. Do the baselines use comparable terminal values and policy proposals, and what is S2P’s total environment-interaction budget? Please offer a matched-data comparison, including success without injected delays, separate control-learning differences from the effect of planning speed.

  2. Is encoder noise sampled during training, or is Eq. (6) evaluated only at encoder means? What implementation constraint or evidence rules out latent rescaling as the explanation for the norm penalty, and supports the claimed information-removal mechanism?

  3. Does S2P outperform an identical pooled-DINO/MLP controller trained with conventional reward/dynamics supervision when data, value learning, planning, and optimization budgets are matched, with repeated training seeds and independently evaluated planner selection?

3 comments
AuthorsOpenReview ↗

Response to Reviewer fuu8

We thank Reviewer fuu8 for their time and helpful feedback, in particular for the insight into the limitations of our deterministic simplification of the BIB loss. In our updated revision, we have matched training protocols, value learning and policy proposals between S2P and baselines, and implemented S2P as a fully stochastic model, with the full variational BIB loss. We also provide evidence that S2P’s BIB loss filters irrelevant information, via a probing analysis on a distracting task.

Addressing Weaknesses

“S2P trains online, while the baselines train offline … S2P also explicitly includes TD-MPC2 value learning and policy-proposed trajectories, whereas the baseline adaptation [does not].”

To address the training-protocol confound in the comparison to DINO-based methods, we train S2P and all DINO-based baselines online, under matching training protocols. In addition, we provide the baselines with value & policy heads to match S2P. Under equivalent architectures and training protocols, S2P retains its success rate advantage against the DINO-based baselines (Table 1).

“Appendix A.1 provides a valid stochastic variational lower bound, but fixed nonzero Gaussian covariance still requires expectations over encoder noise”

We thank the reviewer for their insight into how a rescaling of the latent can minimize the deterministic form of the BIB objective (original submission, Eq. 6). To account for this, we have updated our S2P implementation, to directly optimize the variational lower bound in Eq. 4 using a stochastic encoder.

“The contribution of the proposed objective beyond established compact world-model learning is insufficient. For example, DeepMDP … [and] TD-MPC2”

To further support our contribution of the BIB as a compressive loss, we compare our BIB loss against several other losses used to train compact world models. This comparison shows that other losses either fail to train the S2P architecture (Fig. 5b), or, in the case of TD-MPC2, retain more distractor information (Fig. 6b,c).

“The reported uncertainty is insufficient for the smaller performance and ablation claims”

To better characterize the variance across training seeds, we train S2P, BIB ablations, and the alternate compression losses under three different training seeds. In addition, we evaluate S2P and the DINO-based baselines over 25 episodes per planner configuration, and the ablations and alternate losses over 50. Due to their high online training cost (Sec. 5.3), we keep one training seed for the DINO-based baselines. Our results remain consistent when adding these additional training seeds and evaluation episodes (Table 1, Fig. 4).

Responses to Questions

“Do the baselines use comparable terminal values and policy proposals, and what is S2P’s total environment-interaction budget? Please offer a matched-data comparison, … [and] success without injected delays”

Yes – in our revised comparison, we provide the baselines with a value head and policy proposal, matching S2P’s, and train all methods online with matching training protocols (1M gradient descent steps, one transition collected per step). We also report results without simulated latency effects. Our results show that S2P maintains its success rate advantage under this matched evaluation (Table 1, Fig. 4).

“Is encoder noise sampled during training? … What implementation constraint or evidence rules out latent rescaling … and supports the claimed information-removal mechanism?”

Our updated method samples encoder noise during training, directly optimizing our variational objective with stochastic models. Due to the now-active entropy term, shrinking the latent no longer reduces the loss, and removing the entropy term, which reintroduces this degeneracy, causes training to fail (Fig. 5a).

We additionally provide a probing analysis on S2P latents, trained over a distracting task. Our probes (Section A.6) show that the BIB loss removes task-irrelevant information better than a TD-MPC2 loss does (Fig. 6b,c).

“Does S2P outperform an identical pooled-DINO/MLP controller trained with conventional reward/dynamics supervision when data, value learning, planning, and optimization budgets are matched, with repeated training seeds and independently evaluated planner selection?”

Yes – our updated submission now includes a multi-seed comparison of training S2P with our BIB loss, vs other losses. We train each loss online, under matched training protocols and architectures. This comparison shows that other losses either fail to train the S2P architecture (Fig. 5b), or, in the case of TD-MPC2, retain more distractor information (Fig. 6b,c).

To avoid planner-selection bias, Fig. 5 reports the full distribution over all 54 planner configurations rather than the best one, and Table 1b uses a single fixed configuration.

Thanks for the reponse and experiments. I still have several questions:

  1. S2P uses a planner to collect training data, while the baselines use only their policy networks. Why is there a difference, is it fair?

  2. What benefit does the KL term bring? Now, removing the KL term gives similar or even better results.

  3. The baselines still use only one training seed. Can you use more seeds?

Thank you for reviewing our newest results and experiments.

“S2P uses a planner to collect training data, while the baselines use only their policy networks. Why is there a difference, is it fair?”

We acknowledge this asymmetry. When adapting the baselines to online training, we found that collecting training data with planning would make each baseline run take ~ 27 days on an NVIDIA H100 (because planning with baselines is 1–2 orders of magnitude slower than S2P, Fig. 3). At that point, re-running all our S2P experiments under policy-head collection was not feasible within the rebuttal timeline. We acknowledge that training S2P using the policy to collect data would strengthen the comparison, and will train S2P with this method for the final version of our paper.

We do not expect that this change will substantially alter our results, and we can offer some preliminary evidence for this. We trained S2P with policy-head collection on LiftPeg, PushCube, and PickCube (two seeds, still running), evaluating periodically with the same MPPI planner used throughout our main S2P runs. On PickCube, the evaluation returns and success rates of policy-collected and planner-collected S2P closely overlap through ~500K gradient steps, reaching ~30-40 (a fully trained S2P agent reaches ~40). In PushCube and LiftPeg they appear to trend similarly, or show policy-collection potentially ahead on LiftPeg. However both PushCube and LiftPeg are earlier in training at ~250K steps, so the comparison is not yet as informative. These results suggest that, at least during early training, S2P learns comparably under both data-collection methods.

To view the training curves, please see policy_acting_vs_planning_return_2x3.png in our supplementary material.

“What benefit does the KL term bring? Now, removing the KL term gives similar or even better results.”

We agree that, on LiftPeg and Distracting LiftPeg, removing the KL term results in similar, or possibly improved success rate (Fig. 5a, Fig. 6a). However, the presence of the KL term provides a mechanism to tune the degree of compression by adjusting the value of β\beta. Our results on distracted LiftPeg suggest that β=0.5\beta=0.5 may have some small, but consistent effect on decreasing the amount of distracting information contained in the S2P latent. Below, we tabulate the per-seed probe accuracies and variance shares (i.e. bar heights in Fig. 6b,c).

SeedProbe Accuracy (%): S2PProbe Accuracy (%): S2P-NoKLVariance Shares (%): S2PVariance Shares (%): S2P-NoKL
125.9728.430.4560.532
226.1726.670.5580.447
325.6328.190.5890.572
426.2628.730.5180.650

In probe accuracy, we see that all S2P (β=0.5\beta=0.5) seeds are below all seeds of S2P-NoKL. Variance shares do not show as clear of a trend. While the magnitude of the effect is small, we suspect that it may become more pronounced in tasks with more distractors, which we plan to explore in future work.

“The baselines still use only one training seed. Can you use more seeds?”

Running multiple seeds for our DINO-based baselines is not feasible within the timeline of the rebuttal period; running one seed of all baselines requires 757 hours (~31 days) on NVIDIA H100 GPUs. We have updated our discussion in Sec. 5.3 to explicitly mention that this is the reason for running just one seed of the baselines. The original papers for our baselines likewise do not specify training seed variability (Zhou et al. 2025; Chun et al. 2026; Fu et al. 2026; Toso et al. 2026).

Reviewer_2WAnOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The author points out that current world models typically perform dynamic predictions in high-dimensional visual feature spaces, resulting in high computational costs during the planning process and making it difficult to meet the requirements of real-time robot control. The authors aim to compress the frozen DINOv3 features into a compact, task-relevant state representation through reward prediction, dynamics prediction, and latent state regularization. Using this representation, the authors use a lightweight MLP to build the dynamics and reward models, then combine them with MPPI for real-time action planning. I believe this paper addresses a real-world problem and achieves a notable improvement in computational efficiency. This is also a meaningful area of research that deserves our attention.

Critical strengths

As visual foundation models are increasingly applied to robotics, how to reduce the computational cost of model-based planning while preserving the capabilities of pre-trained visual representations has become a topic worthy of research. This paper addresses a valuable and novel problem: the trade-off between high-dimensional visual representations and real-time planning efficiency, and explores efficient planning through task-specific representation compression.

  1. The authors propose a concise and effective framework for learning compact visual representations. This design enables the model to perform dynamics prediction using a lightweight MLP, thereby avoiding planning directly in the high-dimensional visual token space. While the individual modules are not complex in and of themselves, combining pre-trained visual representations, task-specific compression, and efficient model predictive control constitutes a sound and practical design.

  2. The experimental results in this article are also very impressive. Figure 3 illustrates the significant advantages of S2P over existing DINO-based methods in terms of planning latency and GPU memory usage. More importantly, Figure 4 and Table 2 further demonstrate that these efficiency gains translate into higher task success rates under real-time constraints, rather than merely reducing computational costs.

  3. This paper investigates whether the robustness of pre-trained models can be preserved even after significantly compressing visual representations by introducing perturbations in lighting conditions and object colors. It demonstrates that compact, task-relevant state representations do not necessarily conflict with the robustness of pre-trained visual models.

Critical weaknesses
  1. Table 3 shows that removing the Dynamics Loss from the encoder’s gradients results in a slight increase in task success rates. This observation contrasts somewhat with the authors’ theoretical rationale, which emphasizes effective state compression through the preservation of dynamic information. Therefore, I believe this paper requires more direct comparative experiments. For example, I suggest that the authors compare different training objectives within the same compression architecture to demonstrate the necessity of the Bisimulation Information Bottleneck and its independent contribution.

  2. Second, Figure 4 and Table 2 show that S2P significantly outperforms other DINO-based methods in real-time operation tasks. However, these performance improvements may stem from both lower planning latency and more efficient state representations. In particular, for tasks with strict time constraints, some baselines may be unable to execute a sufficient number of actions due to high computational overhead. Their lower success rates do not necessarily reflect shortcomings in the underlying state representations themselves. Supplementary performance comparisons in environments without real-time latency constraints would help further clarify the contributions of this paper.

Questions for authors

Table 3 shows that removing Dynamics Loss from the encoder’s gradients does not reduce the task success rate. Could the authors explain this phenomenon further?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 2WAn

We thank Reviewer 2WAn for their time and helpful feedback. In our updated submission, we have addressed the weaknesses by providing additional evidence to support the necessity of S2P’s BIB training loss against other pre-existing compressive losses, and BIB ablations. Furthermore, we include performance on environments without real-time latency constraints, to show that S2P’s low-dimensional representations improve task performance independently of planning latency.

Addressing Weaknesses

“Removing the Dynamics Loss from the encoder’s gradients results in a slight increase in task success rates”

In the updated revision of our paper, we have improved S2P’s theoretical grounding in variational information bottlenecks, by directly implementing Eq. 4 with a fully stochastic model. With this updated modelling choice, we see that the dynamics, reward, and entropy components of the BIB loss are each critical for training the S2P architecture (Fig. 5a). Furthermore, both the full and No-KL variants of the BIB loss show a ≥ 8 percentage-point success rate improvement over the TD-MPC2 loss on default LiftPeg and on our new distracted LiftPeg task (Fig. 5b, Fig. 6a). Additionally, they both show near-perfect erasure of task-irrelevant information about distracting particles in our new distracted LiftPeg task (Fig. 6b,c).

“Compare different training objectives within the same compression architecture”

To further support our contribution of the BIB as a compressive loss, we compare our BIB against several other losses used to train compact world models. This comparison shows that other losses either fail to train the S2P architecture (Fig. 5b), or, in the case of TD-MPC2, retain over 5× more distractor information distractor information (Fig. 6c).

“S2P significantly outperforms other DINO-based methods in real-time operation tasks … Supplementary performance comparisons in environments without real-time latency constraints would help further clarify”

To disambiguate whether S2P’s performance gain stems from better representation efficiency, or simply a reduced planning latency, we include results for S2P and baselines on environments without simulated latency effects (Table 1b). Even without the additional latency, S2P matches or outperforms the DINO-based methods.

Responses to Questions

“Removing Dynamics Loss from the encoder’s gradients does not reduce the task success rate. Could the authors explain … ?”

Our initial results, using the deterministic form of S2P, showed that preventing gradients from S2P’s dynamics loss from reaching its encoder resulted in a slight increase in the mean success rate on LiftPeg. After switching to the more theoretically grounded stochastic form of S2P, this is no longer the case. Under the updated loss, the objective can no longer be reduced by rescaling the latent. We see that the dynamics, reward, and entropy components of the BIB loss are each critical for training the S2P architecture. Without dynamics or entropy gradients, training fails entirely (0.00 median success across three seeds, Fig. 5a). Furthermore, both the full and No-KL variants of the BIB loss show an ≥ 8 percentage-point success rate improvement over the TD-MPC2 loss on the default and new distracted version of our LiftPeg task (Fig. 5b, Fig. 6a). Additionally, they both show near-perfect erasure of task-irrelevant information about distracting particles in our distracted LiftPeg task (Fig. 6b,c).

As Reviewer fuu8 also pointed out, I am still unclear about the action collection process. Do the baselines also use MPPI for online data collection, as S2P does, or do they collect data using their own policies? Different collection strategies may lead to different state distributions. Could the authors clarify this?

We thank you for reviewing our comments and updated results.

"Do the baselines also use MPPI for online data collection, as S2P does, or do they collect data using their own policies?"

We acknowledge this asymmetry. When adapting the baselines to online training, we found that collecting training data with planning would make each baseline run take ~ 27 days on an NVIDIA H100 (because planning with baselines is 1–2 orders of magnitude slower than S2P, Fig. 3). At that point, re-running all our S2P experiments under policy-head collection was not feasible within the rebuttal timeline. We acknowledge that training S2P using the policy to collect data would strengthen the comparison, and will train S2P with this method for the final version of our paper.

We do not expect that this change will substantially alter our results, and we can offer some preliminary evidence for this. We trained S2P with policy-head collection on LiftPeg, PushCube, and PickCube (two seeds, still running), evaluating periodically with the same MPPI planner used throughout our main S2P runs. On PickCube, the evaluation returns and success rates of policy-collected and planner-collected S2P closely overlap through ~500K gradient steps, reaching ~30-40 (a fully trained S2P agent reaches ~40). In PushCube and LiftPeg they appear to trend similarly, with policy collection possibly ahead on LiftPeg. However both PushCube and LiftPeg are earlier in training at ~250K steps, so the comparison is not yet as informative. These results suggest that, at least during early training, S2P learns comparably under both data-collection methods.

To view the training curves, please see policy_acting_vs_planning_return_2x3_full_1M.png in our supplementary material.

Reviewer_Ecv5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

S2P uses cross-attention pooling to compress frozen DINOv3 features and proprioception into a 512-dimensional state. MLP dynamics and reward models, together with value learning, support MPPI planning in this space. The objective is motivated by bisimulation and information bottlenecks and implemented through reward prediction, dynamics prediction, and latent L2 regularization. Experiments on three modified ManiSkill3 tasks show lower planning latency and memory use, improved success under simulated planning delays, and robustness to several lighting and colour changes.

Critical strengths
  • The method targets the cost of repeatedly evaluating dynamics during planning. Compressing the state into a single vector allows a lightweight MLP world model. Under the same planner configuration, Figure 3 reports 161 MiB peak memory for S2P versus 7.51 GiB for DINO-WM, with planning speedups of roughly one to two orders of magnitude. This is a clear and practically useful gain.
  • The method reduces planning cost while retaining robustness to several appearance changes. S2P performs consistently under changes in brightness, colour temperature, and object colour, whereas the encoder trained from scratch degrades sharply. This supports retaining pretrained visual priors rather than pursuing efficiency alone.
Critical weaknesses
  • The success-rate comparison does not separate representation learning from the training protocol. S2P collects data and updates online, while all DINO baselines use fixed offline datasets. Their state distributions and opportunities to adapt to their own planning errors therefore differ. The computational advantage remains clear, but attributing the success-rate gap to the compression method requires further evidence.
  • The benefit of dynamics supervision for learning the compressed representation is unclear. In Table 3, blocking dynamics-loss gradients to the encoder changes success from 63% to 67%, without degradation. The ablation therefore does not demonstrate its contribution to learning task-relevant representations.
Questions for authors

Does S2P retain its success-rate advantage when trained on the same offline datasets as the baselines? This would help separate the benefit of compression from that of online data collection.

1 comment
AuthorsOpenReview ↗

Response to Reviewer Ecv5

We thank Reviewer Ecv5 for their time and helpful feedback, particularly in strengthening our comparison against the baselines. Our updated submission addresses the stated weaknesses, by addressing the training-protocol confound in our comparisons against the DINO baselines, and providing evidence to support the necessity of the BIB loss in training S2P.

Addressing Weaknesses

“Benefit of dynamics supervision…unclear.”

In our updated submission, we have improved S2P’s theoretical grounding in variational information bottlenecks by directly implementing Eq. 4 with a fully stochastic model. With this updated modelling choice, we see that the dynamics, reward, and entropy components of the BIB loss are each critical for training the S2P architecture (Fig. 5a). Furthermore, both the full and No-KL variants of the BIB loss show an ≥ 8 percentage-point success rate improvement over the TD-MPC2 loss on the default and distracted version (new) of our LiftPeg task (Fig. 5b, Fig. 6a). Additionally, they both show near-perfect erasure of task-irrelevant information about distracting particles in our new distracted LiftPeg task (Fig. 6b,c).

“Comparison does not separate representation learning from the training protocol…computational advantage remains clear, but attributing the success-rate gap to the compression method requires further evidence.

To address the training-protocol confound in the comparison between S2P and DINO-based baselines, we train all DINO-based baselines online, under matching training protocols. In addition, we further augment the baselines by adding learned value and policy heads to all DINO-based world models. These new results show that under matched training protocol, value learning, and policy proposals, S2P retains its success rate advantage against these methods (Table 1, Fig. 4).

Responses to Questions

“Does S2P retain its success-rate advantage when trained on the same offline datasets as the baselines?”

No – despite S2P’s success in online training, we are currently unable to achieve the same success when training offline, despite trying a variety of datasets. To keep all training protocols consistent, our updated submission trains all methods online, with identical environment interaction budgets.

Given that S2P learns a minimal sufficient statistic to predict latent dynamics and reward over the support of the dataset, we postulate that this statistic is sensitive to distributional differences in the offline dataset vs the online visitation. We have clarified our discussion of this limitation in Sec. 4.1, and highlight this exciting avenue for future work in our Conclusion.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We thank all reviewers for their thoughtful feedback. Major changes in the revised PDF are highlighted in blue. In summary:

  1. S2P is fully stochastic: We replace the deterministic simplification of the BIB loss with a fully stochastic model, directly optimizing the variational objective (Eq. 4) with learned-variance Gaussians (Sec. 4.2). Under this objective, shrinking the latent no longer reduces the loss.

  2. Matched online training protocols. All methods, including the DINO-based baselines, are now trained online for 1M gradient steps with matched update-to-data ratios. Baselines are also augmented with value and policy heads matching S2P's (Sec. 5.3). Under this matched comparison, S2P retains its success-rate advantage (Table 1a, Fig. 4).

  3. Evaluation without real-time latency: We report success rates without simulated planning delays, using a single fixed planner configuration (Table 1b), to disentangle the effects of representation quality versus planning speed.

  4. Loss ablations and alternate losses: We ablate each BIB term, and compare against training with TD-MPC2, DeepMDP, and RePo losses on the S2P architecture. We report results over three training seeds, and the full distribution of success rate across 54 planner configurations (Fig. 5).

  5. Probing analysis for task-irrelevant information: On a new distracted LiftPeg task, we show that BIB-trained latents retain substantially less distractor information than TD-MPC2-trained latents, as demonstrated by a probe analysis. BIB-trained latents also show an over 5$\times$ reduction of latent variance attributable to the distractor compared to TD-MPC2-trained latents (Fig. 6, App. A.6).

We address each reviewer's specific concerns, and restate the relevant changes in our individual responses below.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.