acceptodds
Under review as a conference paper at ICLR 2027

Embodied Learning in the Loop: Co-Adapting Policies and Environment Distributions

Abstract

We devise a learner-aware framework for embodied PointNav that treats the scene distribution as a policy-conditioned component of training. Existing policies are trained on static corpora or samples from pretrained generators, neither of which responds to the learner’s evolving competence. Moreover, scene-centric properties (e.g., realism, reachability, geometric complexity) do not identify scenes that promote learning. Our framework adapts a layout diffusion generator using scene-level policy learning progress. It generates indoor layouts for PointNav and compares policy performance on the same episodes before and after each policy update. Changes in success, path efficiency, and goal distance define a utility signal that weights the denoising records that produced each scene, steering updates to the pretrained generator. This closes the loop between scene generation and policy learning. Averaged across five distributions, our pipeline improves PointNav performance (5.84 and 4.44 percentage-point gains in SR and SPL, respectively, over Frozen Generator). Policy feedback shifts the generator to a distinct distribution (MMD²=0.752), with the separation persisting across later adaptation intervals. Matched-success trajectories and same-input comparisons further reveal distinct navigation behaviors and action choices under identical observation and action histories. Our code will be released.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Embodied Navigation, Adaptive Environment Generation, Learner-Aware Training

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_U1eTOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper jointly adapts a scene generator and a PointNav policy. Policy improvement on fixed before/after episodes provides scene-level weights for updating a diffusion generator through LoRA. Experiments evaluate navigation across five distributions and analyze changes in generated layouts and policy behavior.

Critical strengths
  1. The external-domain results are promising. In Table 1, MP3D success improves from 73.94% to 83.64%, and SPL from 51.38% to 59.79%, over Frozen Generator with the same policy training budget. This supports the practical value of adapting training environments.

  2. Table 2 includes selection-only and shuffled-feedback controls. Full performs better on both external domains, providing useful evidence that the scene-specific feedback matters.

Critical weaknesses
  1. Tables 1–2 do not report the number of independent training runs or variation across seeds. Both policy learning and scene generation are stochastic, so I am unsure how reliable the gains are. The bootstrap analyses elsewhere do not measure training-run variation.

  2. The central generator update needs more detail. Eqs. 10–11 do not clearly explain how the target noise is constructed from the retained generation record. The utility weights and clipping settings are also missing, making the method difficult to reproduce.

Questions for authors
  1. Clarification and supporting evidence: How many independent training seeds were used for Tables 1–2? Can you report mean and standard deviation for Full, Frozen Generator, and Shuffled-feedback on Gibson and MP3D?

  2. Clarification: How are the noisy state and target noise in Eqs. 10–11 constructed? Are generated layouts forward-noised, or are reverse-sampling states reused? Please also specify the utility weights and clipping settings.

2 comments
AuthorsOpenReview ↗

Point-to-Point Response to Reviewer U1eT

We thank reviewer U1eT for the valuable time and constructive feedback. We provide point-to-point responses below.

Q1: Additional training seeds and result variability.

A1: The results in Tables 1–2 were obtained from one training run per method. We conduct two additional independent runs for Full, Frozen Generator, and Shuffled-feedback, bringing the total to three training seeds per method. The table below reports mean ± sample standard deviation across training runs on Gibson and MP3D.

MethodGibson SR (%)Gibson SPL (%)MP3D SR (%)MP3D SPL (%)
Frozen Generator91.04 ± 0.6169.47 ± 1.5176.23 ± 2.0253.73 ± 2.39
Shuffled-feedback91.68 ± 0.7460.16 ± 0.5178.05 ± 1.6550.97 ± 2.40
Full94.27 ± 0.4670.49 ± 0.5983.71 ± 1.1159.87 ± 0.27

Full achieves the highest mean SR and SPL on both external domains, with lower standard deviations than both controls across all four metrics. These results support the consistency of the gains across independent training runs. We have added the results to supp. §C.2, Table 11.

Q2: Clarification on generator updates, utility weights, and clipping settings.

A2: We forward-noise the completed generated layouts using freshly sampled noise, which serves as the denoising target. Reverse-sampling states are not reused. The SR, SPL, and normalized-DTG coefficients are 1, 0.5, and 0.1, respectively. We detail the construction and weighting settings below.

First, generated layouts are forward-noised; intermediate reverse-sampling states are not reused. We retain each completed generated layout, x0x_0, and its condition, cc. At each adaptation update, we sample a timestep tt and fresh standard Gaussian noise ϵ∼N(0,I)\epsilon\sim\mathcal{N}(0,I), and construct the noisy input as

xt=αˉtx0+1−αˉtϵx_t=\sqrt{\bar{\alpha}_t}x_0+\sqrt{1-\bar{\alpha}_t}\epsilon.

Here, αˉt\bar{\alpha}_t is the cumulative signal-retention coefficient at timestep tt. The sampled ϵ\epsilon is the target noise. The network's vv-prediction is converted to a noise prediction before computing the feedback-weighted noise MSE. We have clarified this construction and the retained records in §3.5 and supp. §A.4.

Second, we specify the utility weights and clipping settings. The navigation score is

Qi=SRi+0.5SPLi−0.1clip⁡(DTG‾i,0,10)10Q_i=\mathrm{SR}_i+0.5\mathrm{SPL}_i-0.1\frac{\operatorname{clip}(\overline{\mathrm{DTG}}_i,0,10)}{10}.

SR and SPL are in [0,1][0,1]. The term DTG‾i\overline{\mathrm{DTG}}_i is the mean distance-to-goal in meters across the scene's evaluation episodes. The function clip(value, lower, upper) limits the value to the specified interval. Feedback and weights are computed in the following order:

  1. Learning progress: ri=clip⁡(Qafter,i−Qbefore,i,−3,3)r_i=\operatorname{clip}(Q_{\mathrm{after},i}-Q_{\mathrm{before},i},-3,3).
  2. Exponential weighting: ui=exp⁡(clip⁡(10ri,−6,6))u_i=\exp(\operatorname{clip}(10r_i,-6,6)).
  3. Normalization and clipping: wi=clip⁡(ui1∣P∣∑j∈Puj,0.05,4)w_i=\operatorname{clip}(\frac{u_i}{\frac{1}{|\mathcal{P}|}\sum_{j\in\mathcal{P}}u_j},0.05,4),

where P\mathcal{P} is the eligible fitting pool, and the denominator in the final expression is the mean of uju_j over that pool. Normalization precedes the final weight clipping, so the final weights need not have exactly unit mean.

The coefficients were selected on the ProcTHOR validation set. The selection procedure and sensitivity results are provided in supp. §C.4. We have also specified the feedback-weighting settings and operation order in §3.5, supp. §A.4, and Table 8.

Thanks for the rebuttal. I think the three-seed results and the clarified generator update address both of my concerns. I think I would raise my rating to 3.

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

This paper proposes a framework that jointly adapts 3D scene generation and navigation policy learning using policy feedback. The authors use a diffusion-based model (DiffuScene) to generate indoor object layouts, which are converted into navigation environments using ProcTHOR assets and Habitat. A PointNav policy is trained using PPO, and its learning progress is measured by comparing navigation performance on the same evaluation episodes before and after each policy update. This feedback is then used to update the scene generator through a weighted diffusion loss with LoRA, allowing the generated scene distribution to change during training. The method is evaluated on five indoor scene distributions, including Gibson and Matterport3D, and further analyzed through generator distribution changes and navigation behavior comparisons.

Critical strengths
  1. The idea of co-adapting scene generation and navigation policy learning is interesting and provides a practical way to improve navigation training.
  2. Using policy learning progress as feedback to update the scene generator is a reasonable approach that allows the training environments to adapt as the policy improves.
  3. The method achieves promising navigation success rates across five evaluation distributions, including unseen Gibson and Matterport3D environments.
Critical weaknesses

The learning-progress feedback may favor scenes where the policy improves quickly, while giving less importance to difficult scenes that could be useful for learning.

It is unclear how the authors chose the weights in the scene utility equation, and the paper does not study how different weights affect the results.

Although the method improves generalization, the paper does not clearly explain what changes in the generated scenes lead to better navigation performance.

The paper measures learning progress using only eight evaluation episodes per scene. The paper does not examine whether this small number of episodes provides reliable feedback.

The paper does not compare its learning-progress feedback with other approaches, such as difficulty-based, regret-based, or navigation-score-based feedback, making it unclear whether the proposed approach is more effective.

Questions for authors

Related to the weakness that I think should be improved:

(1) It is unclear how the authors chose the weights in the scene utility equation, and the paper does not study how different weights affect the results.

(2) The paper does not clearly explain what changes in the generated scenes lead to better navigation performance.

6 comments
AuthorsOpenReview ↗

Point-to-Point Response to Reviewer yYFQ (1/2)

We thank reviewer yYFQ for the valuable time and constructive feedback. We provide point-to-point responses below.

Q1: Discussion on learning progress and difficult scenes.

A1: Difficulty alone does not establish training utility. A scene that offers little immediate progress may nevertheless become useful later. We distinguish this concern about delayed utility from the overall effectiveness of the feedback signal.

First, scene utility is recomputed as the learner changes. Our method does not assign difficult scenes a permanently low utility. A scene can receive greater weight when subsequently sampled at a stage where the policy can learn from it. Thus, the feedback responds to the learner’s changing competence rather than assigning difficult scenes a fixed low priority.

Second, the difficulty-based comparison tests overall performance, rather than coverage of all useful difficult scenes. Difficulty-based feedback uses 1−SRafter1-\mathrm{SR}_{\mathrm{after}}, assigning higher feedback to scenes with lower post-update success rates. Across three training seeds, learning progress improves Gibson SR/SPL by 3.76/7.96 percentage points and MP3D SR/SPL by 4.86/8.01 percentage points (supp. §C.5, Table 14). These results support measuring learning progress rather than difficulty alone when guiding scene adaptation.

Q2: Selection and sensitivity of utility weights.

A2: The utility coefficients were selected through short pilot experiments on the ProcTHOR validation set. We fix the SR coefficient to 1 and vary the SPL and normalized-DTG coefficients, selecting 0.5 and 0.1, respectively.

To examine sensitivity, we compare five configurations from the same checkpoint. Each completes five effective generator updates and is evaluated on the same 500 ProcTHOR validation episodes. The variants halve or double the SPL coefficient, remove the DTG term, or double its coefficient.

ConfigurationSPL weightDTG weightSR (%) ↑SPL ↑
Shared starting checkpoint——36.400.2100
Selected weights0.500.1068.000.3882
Halved SPL weight0.250.1059.000.3350
Doubled SPL weight1.000.1064.400.3727
Without DTG0.500.0057.000.3411
Doubled DTG weight0.500.2061.400.3469

The selected weights achieve the highest validation SR and SPL among the tested configurations. Both halving and doubling the SPL coefficient reduce performance, while removing the DTG term also lowers SR and SPL. Thus, the coefficients affect short-run adaptation performance, and the validation comparison supports the selected setting. We have added the selection procedure and sensitivity results to supp. §C.4, Table 13.

AuthorsOpenReview ↗

Point-to-Point Response to Reviewer yYFQ (2/2)

Q3: Analysis of scene changes and navigation improvements.

A3: Our analyses show that feedback-adapted layouts are more navigable, with higher start–goal reachability, lower path detour, and greater clearance, while the resulting policy exhibits stronger local recovery. We detail the scene-level and policy-level evidence below.

First, feedback-adapted layouts exhibit improved navigability. We add the following scene-level comparison between Frozen and feedback-adapted layouts (supp. §C.8, Table 17).

MetricFrozen generatorFeedback generator
Robot-height obstacle occupancy proxy11.00%5.34%
Retained traversable-area fraction83.90%91.60%
Reachable start–goal proposal fraction69.06%83.54%
Shortest-path / Euclidean distance1.2791.218
Path-clearance P10 (m)0.1050.149
Within-path fraction of samples with clearance < 0.2 m33.23%23.36%

The adapted layouts have higher start–goal reachability (69.06% → 83.54%), lower path detour (1.279 → 1.218), and greater path clearance (P10: 0.105 m → 0.149 m). These results characterize a more traversable training distribution. The within-condition analysis additionally shows increased spatial variation (supp. §C.7).

Clearance is measured relative to the boundary of the robot-eroded navmesh and therefore represents additional clearance within the navigable region, rather than physical passage width. Detailed sampling and aggregation procedures are provided in supp. §C.8.

Second, the resulting policy exhibits stronger local recovery. Pooling paired continuations from matched stalled states collected under both policies, Full increases local-escape rates from 43.3% to 80.0% on Gibson and 37.5% to 61.7% on MP3D, while reducing invalid-forward rates on both domains (§4.5, Table 7; supp. §C.11, Table 19). This comparison shows improved recovery under matched starting conditions.

Together, these analyses characterize a more navigable and spatially varied training distribution, accompanied by stronger policy recovery. The scene descriptors are used only for post-hoc analysis, not for feedback-based scene selection, and do not individually establish which layout changes cause the navigation gains.

Q4: Reliability of feedback from eight evaluation episodes.

A4: Scene-specific feedback computed from eight paired episodes provides useful guidance for generator adaptation. This budget balances online evaluation cost with the practical utility of the feedback.

Each iteration requires 16 evaluation rollouts per scene, with the same eight episodes evaluated before and after the policy update. Reusing the episodes removes variation from resampling different episodes, although finite-sample noise remains.

To examine whether the resulting feedback is informative, Shuffled-feedback preserves each iteration’s feedback values but permutes their scene assignments. Full outperforms this control under the same feedback budget (Table 2), and the added three-seed comparison shows higher mean SR and SPL on both Gibson and MP3D (supp. §C.2, Table 11). These results support the value of the scene–feedback associations despite sampling noise, as also recognized by Reviewer U1eT.

We have clarified the budget rationale in supp. §B.2. This comparison establishes the practical utility of feedback computed from eight paired episodes; it does not establish that eight is optimal or that the resulting scene-level LP estimates are precise.

Q5: Comparisons with alternative feedback signals.

A5: Learning-progress feedback achieves higher mean SR and SPL than all three tested alternatives on both external domains. We add comparisons with difficulty-based, navigation-score-based, and regret-based feedback (supp. §C.5, Table 14), reporting mean ± sample standard deviation across three independent training seeds. All metrics are expressed as percentages.

Feedback signalGibson SR (%)Gibson SPL (%)MP3D SR (%)MP3D SPL (%)
Difficulty90.51 ± 0.5762.53 ± 2.1278.85 ± 1.5451.86 ± 2.99
Navigation score93.09 ± 1.4168.13 ± 3.2179.93 ± 4.5054.36 ± 3.14
Regret92.29 ± 1.2560.10 ± 2.3881.08 ± 2.2650.64 ± 1.86
Learning progress (Full)94.27 ± 0.4670.49 ± 0.5983.71 ± 1.1159.87 ± 0.27

Learning-progress feedback achieves the highest mean SR and SPL on both external domains, with the lowest standard deviation across training seeds on all four metrics. Compared with navigation-score feedback, it improves Gibson SR/SPL by 1.18/2.36 percentage points and MP3D SR/SPL by 3.78/5.51 percentage points. These results support using measured policy improvement to guide scene adaptation.

I thank the authors for their detailed response and additional experiments. The new sensitivity analysis, alternative feedback comparisons, and scene-level analysis address most of my concerns and strengthen the paper.

" Difficulty alone does not establish training utility. A scene that offers little immediate progress may nevertheless become useful later. We distinguish this concern about delayed utility from the overall effectiveness of the feedback signal." -> I do understand this intuitively. Could you possibly provide either experimental evidence or theoretical analysis/proof to support this?

Second, the shuffled-feedback experiment supports the usefulness of the feedback signal, but it does not directly establish how reliable the learning-progress estimates are when computed from only eight episodes.

Overall, I appreciate the authors' efforts and the additional experimental evidence.

Reviewer_Ry44OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

It is a method that closes a loop between a layout diffusion generator and a PointNav policy. DiffuScene trained on ProcTHOR generates scenes that are instantiated in Habitat. A PPO policy trains on them. The change in navigation score weights the denoising loss to adapt the generator with LoRA. Experiments on five evaluation domains report higher average SR and SPL than a frozen generator, a shift in the generated distribution, and different navigation behavior between the two policies.

Critical strengths
  1. Scene usefulness is measured directly as the policy’s improvement on the same episodes before and after each PPO update (sec. 3.4) instead of using properties of the scene itself. This might be the right quantity to use for adapting the generator as discussed in sec 1, which I find interesting.

  2. The MMD**2 and classifier AUC results in table 3 show that the feedback adapted generator produces a very different distribution from the frozen one.

Critical weaknesses
  1. Gains are not consistent. On MP3D, ‘Ours Full’ improves both SR and SPL over the ProcTHOR trained policy. On Gibson it improves SR by ~3.2 points but has lower SPL than ProcTHOR, 0.709 vs 0.731, table 1. The generalization claim therefore rests mainly on MP3D. I would suggest the authors run more rigorous experiments, at least several seeds and one more external domain, before making this claim.

  2. Limited validation. Everything is tested with one generator (DiffuScene), one scene source (ProcTHOR), one task (PointNav) and one policy. Running the loop with at least one other layout generator would show the feedback adaptation is not specific to DiffuScene.

  3. No human validation of the generated scenes. sec. 4.3 shows the generator distribution moved (table 3) but no figure shows what the new scenes look like. There are no renderings of frozen vs feedback adapted layouts. Also, I recommend authors to have human validation metrics of whether the adapted scenes are still plausible rooms. A small visual comparison or user study on sampled layouts from both generators would show whether the shift is meaningful.

Because of lack of these experiments the claims in the paper aren’t ready for acceptance.

Questions for authors

NA

3 comments
AuthorsOpenReview ↗

Point-to-Point Response to Reviewer Ry44 (1/2)

We thank reviewer Ry44 for the valuable time and constructive feedback.

Clarification of our central claim. We respectfully disagree with the assessment that the missing experiments leave our stated claim unsupported. Our central claim is that adapting a scene generator using policy learning progress improves PointNav generalization, supported by the controlled comparisons in Tables 1–2. Additional training seeds and external-domain evaluations strengthen this evidence, and we provide these results below. However, cross-task effectiveness and human-rated scene plausibility are outside our stated claims. Their absence therefore does not justify the conclusion that our PointNav generalization claim is unsupported.

We then provide point-to-point responses below.

Q1: Consistency of gains and generalization to external domains.

A1: We clarify that our claim concerns improved overall generalization, rather than uniformly higher performance on every metric in every domain. The controlled comparison with Frozen Generator evaluates the benefit of feedback adaptation, while comparison with the ProcTHOR-trained policy provides an additional training-source baseline. Nevertheless, to further address your concern and examine the consistency and generality of the gains, we add independent training runs and a third external evaluation domain, HM3D.

First, the gains persist across independent training seeds. We add two runs each for Full, Frozen Generator, and Shuffled-feedback, bringing the total to three training seeds per method. Results are mean ± sample standard deviation across training runs.

MethodGibson SR (%)Gibson SPL (%)MP3D SR (%)MP3D SPL (%)
Frozen Generator91.04 ± 0.6169.47 ± 1.5176.23 ± 2.0253.73 ± 2.39
Shuffled-feedback91.68 ± 0.7460.16 ± 0.5178.05 ± 1.6550.97 ± 2.40
Full94.27 ± 0.4670.49 ± 0.5983.71 ± 1.1159.87 ± 0.27

Full achieves the highest mean SR and SPL on both external domains, with lower standard deviations than both controls across all four metrics (supp. §C.2, Table 11).

Second, the gains extend to HM3D. Results are mean ± sample standard deviation over three independent training seeds:

PolicySR (%)SPL
Full89.31 ± 0.420.6509 ± 0.0036
Frozen84.33 ± 0.530.6113 ± 0.0020
ProcTHOR-trained86.19 ± 0.510.6314 ± 0.0055
Gibson-trained88.08 ± 0.350.6676 ± 0.0014
MP3D-trained89.25 ± 0.310.6482 ± 0.0007

On HM3D, Full improves SR and SPL over Frozen by 4.98 and 3.96 percentage points, respectively (supp. §C.3, Table 12).

Regarding the SPL comparisons, Full has lower SPL than ProcTHOR-trained on Gibson. Success rate and SPL capture different aspects of navigation performance. SPL combines task completion with path efficiency on successful episodes, whereas our utility jointly weights success, SPL, and remaining goal distance. The results therefore show a remaining path-efficiency trade-off, rather than uniform superiority on every metric. On HM3D, Full improves both SR and SPL over Frozen Generator and ProcTHOR-trained. These comparisons support improved overall generalization without requiring Full to attain the highest SPL in every setting.

AuthorsOpenReview ↗

Point-to-Point Response to Reviewer Ry44 (2/2)

Q2: Validation across generators, scene sources, and tasks.

A2: Our contribution is learning-progress-driven scene adaptation for PointNav. The feedback formulation does not rely on DiffuScene-specific components, the added evaluations test transfer to additional scene settings, and cross-task experiments would extend the stated task scope. We address these points separately below.

First, DiffuScene is our implementation backbone, rather than the basis of a model-specific design. Neither the scene-level learning-progress signal nor its feedback-weighting mechanism uses DiffuScene-specific architectural components. DiffuScene instantiates the proposed loop using a diffusion denoising objective. Evaluating another generator requires a separate integration and retraining pipeline, including layout-representation and asset-category alignment, which cannot be completed within the rebuttal period.

Second, the policy gains transfer to additional scene settings. Adapting a generator trained on another scene source requires data preparation and simulator-compatible asset mapping. To examine transfer beyond the original instantiation setting, we conduct zero-shot evaluations in RoboTHOR and HSSD settings, without retraining either the generators or the policies.

RoboTHOR. We test transfer by instantiating generated layouts with native RoboTHOR structures and assets, using a custom evaluation protocol.

Evaluation layoutsFull SR (%)Frozen policy SR (%)Full SPLFrozen policy SPL
Frozen generator91.783.60.64490.6351
Feedback generator95.189.20.70540.6814

HSSD. We further test the fixed policies on generated layouts instantiated with native HSSD room structures and furniture assets.

Evaluation layoutsFull SR (%)Frozen policy SR (%)Full SPLFrozen policy SPL
Frozen generator90.283.80.63500.6332
Feedback generator91.584.20.67050.6299

Full achieves higher SR and SPL than the Frozen policy on both layout sets in both settings. These results support transfer of the learned policy gains across additional scene settings (supp. §C.1, Table 10).

Third, PointNav is the stated task scope of the paper. The abstract explicitly introduces a framework for embodied PointNav, and supp. §E limits the demonstrated benefits to navigation, identifying interactive tasks as future work. Cross-task experiments would establish an additional capability beyond the PointNav generalization claim evaluated here.

Q3: Visual comparisons and plausibility of generated scenes.

A3: We respectfully disagree that the absence of human validation leaves our generalization claim unsupported. Human-rated plausibility and training utility are different evaluation targets. As emphasized in §1, our objective is to adapt scene generation to improve policy learning. Human evaluation would assess whether the layouts resemble plausible human environments. It would not directly measure whether they improve PointNav learning and transfer.

First, task-oriented evaluation has precedent in environment generation and sim-to-real learning. Domain Randomization demonstrates transfer from simulated images with non-realistic random textures to real robotic tasks (Tobin et al., IROS 2017). PAIRED evaluates generated environments through task complexity and zero-shot policy transfer (Dennis et al., NeurIPS 2020). These works support evaluating training environments through downstream outcomes. They do not establish the human-rated plausibility of our layouts. Accordingly, our evidence for training value comes from controlled policy comparisons and external-domain performance.

Second, we provide direct visual and quantitative inspection of the generated layouts. Fig. 7 in supp. §C.9 compares scene shells, ProcTHOR layouts, and Frozen and feedback-adapted layouts. The navigability analysis reports higher reachability (69.06% → 83.54%), lower path detour, and greater clearance (supp. §C.8, Table 17). These analyses assess task feasibility and spatial properties, while the added training-seed and HM3D results evaluate the resulting policy benefits (supp. §C.2–§C.3). Together, they make the scene changes and their training value explicit. We do not present these measurements as substitutes for human ratings of realism, nor do we claim improved human-rated realism.

Thank you for the detailed response. The three-seed results and HM3D address my concern about consistency. I raise my rating to 3.

One concern still remains. Table 9 shows the pretrained generator places 22% of object centers outside the floor. The gain may therefore come from repairing a weak generator rather than from learner-specific adaptation, which may not appear with a cleaner generator. The RoboTHOR and HSSD results do not test this. They reuse the same generator and policies with different assets.

To rule it out, the method should be run on another generator or scene source as well. I would encourage the authors to add this in the final version.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.