acceptodds
Under review as a conference paper at ICLR 2027

Self-Correcting Imaginative Planning for VLM Embodied Agents

Abstract

Recent advances in Vision-Language Models (VLMs) have demonstrated their remarkable capability to serve as embodied agents in interactive environments. However, these agents often make plans (i.e., action sequences) that appear linguistically plausible but fail upon execution, without considering the visual consequences of each action. While prior works have explored enhancing VLM agents through language supervision or self-correction with text-based verification, enabling them to self-correct invalid plans against visual consequences before execution remains an open challenge. To tackle this, we propose Self-Correcting Imaginative Planning (SCIP), a training framework that enables VLM agents to self-correct their plans against imagined visual consequences before execution. SCIP integrates an imagination-guided policy learning scheme that grounds the VLM agent's planning in visual consequences, and a verification-driven imagination learning scheme that calibrates the imagination process to faithfully reflect the actual environment dynamics. Extensive experiments show that SCIP outperforms existing embodied VLM agents across multiple embodied planning tasks. The code will be released upon acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Vision-Language Model Agents, Embodied Planning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_6cFdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SCIP, a framework that enables VLM agents to revise plans before execution based on predicted visual consequences. In this framework, a action-conditioned world model is introduced to provide visual consequences, and a shared VLM is responsible for planning and verification. The framework is trained with a novel paradigm, imagination-guided policy learning and verification-driven imagination learning. Experiments cover four EmbodiedBench environments, with the strongest improvements observed on EB-Habitat.

Critical strengths
  1. The paper is well-written and easy to follow. The problem is formulated well and the integration of planning, latent prediction, verification, and revision is intuitive and clearly motivated.
  2. The experiments cover both high-level planning and low-level control.
Critical weaknesses
  1. The necessity of two-round alternating training schedule is not well established.
  2. In Table 4, adding Section 3.3 improves success from 57.6% to 64.8%, but this stage bundles action matching, visual similarity, turn-wise improvement, and verifier training. Consequently, the improvement cannot be attributed specifically to imagination-guided learning.
  3. In order to predict the real environment more accurately, the authors introduces verification-driven imagination learning (Sec 3.4). However, Sec 3.4 optimizes predicted latents to increase a frozen verifier’s likelihood of the reference verdict, subject to regularization toward the pretrained model. This may make the latents more useful for classification without necessarily improving their correspondence to actual environment dynamics.
  4. The framework for improving the plan through the prediction of future states have relevant precedents. SCIP's distinction appears to lie primarily in its training paradigm, making the missing ablation studies and analysis particularly important.
Questions for authors

Please see Critical Weaknesses above.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 6cFd (1/2)

(W1) Clarification of the two-stage training schedule and why both stages are needed.

We thank the reviewer for the opportunity to clarify this. The schedule consists of two stages, each run once. Policy learning (Section 3.3) trains the VLM, which acts as both the planner and the verifier, to plan and verify with imagined outcomes, while calibration (Section 3.4) trains the world model to imagine outcomes that lead the verifier to the correct verdict. Specifically, policy learning takes its reward from the imagined outcomes of a fixed world model, while calibration then tunes the world model against the resulting verifier with the VLM fixed. Each stage is thus trained with the output of the other, and each adds its own gain, as the average success rate on EB-Habitat rises from 57.6% (SFT) to 64.8% and then to 68.0% (Table 4), indicating that neither stage is redundant. We have clarified this in Section 3.5 of the revised paper.


(W2) Clarification of how much of the gain of Section 3.3 (57.6% to 64.8%, Table 4) comes from imagination-guided learning.

We thank the reviewer for the opportunity to clarify this. The imagination-guided part of Section 3.3 is the visual-similarity reward (rvisualr_{\text{visual}}, Eq. 2), which credits plans whose imagined final state matches that of the ground-truth plan, whereas action matching and turn-wise improvement compare actions only. To isolate its effect, we additionally retrained Section 3.3 without this reward. As shown in the table below, with all the other components kept, the average success rate drops from 64.8% to 60.8%, indicating that 4.0 of the 7.2 gain of Section 3.3 is specific to the imagination-guided reward. Removing verifier training instead lowers it to 58.8%, so the verifier trained on imagined states also contributes.

Section 3.3 variantEB-Habitat Avg. SRDrop
Full64.8-
Without visual-similarity reward60.8−4.0
Without verifier training58.8−6.0
AuthorsOpenReview ↗

Response to Reviewer 6cFd (2/2)

(W3) Do the imagined states after calibration (Section 3.4) still correspond to the actual environment dynamics?

We thank the reviewer for raising this point. We measure this correspondence by the L1 error between imagined states and real observations in the feature space of the world model (V-JEPA 2 [1]), on the plans that SCIP executes in the EB-Habitat evaluation. As shown in the table below, calibration with verification signals alone raises this error from 0.351 to 0.850 at 4 steps, even though the success rate rises from 64.8% to 68.0% (Table 4). A higher success rate therefore does not ensure this correspondence. To ensure it, we add a next-observation loss to calibration (Section 3.4), which trains the world model to match the real next observation on the training trajectories. With the next-observation loss, the error returns to the level before calibration (0.225 vs. 0.224 at 1 step and 0.358 vs. 0.351 at 4 steps), while most of the gain in success rate is retained (67.7% vs. 68.0%), indicating that the imagined states after the revised calibration still correspond to the actual environment dynamics. We have added this loss to Section 3.4 and this analysis to Section 4.3 (Table 6) of the revised paper.

EB-HabitatL1 ↓, 1 step2 steps3 steps4 stepsAvg. SR ↑
Before calibration0.2240.3550.3450.351–
SCIP (after calibration)0.3510.5680.7060.85068.0
SCIP + next-observation loss0.2250.3550.3480.35867.7

SR averaged over three runs.

[1] Assran et al. V-JEPA 2: Self-supervised video models enable understanding, prediction and planning. arXiv 2025.


(W4) Clarification of what distinguishes SCIP from prior methods that predict future states, and the ablations of its training paradigm.

Thanks to the reviewer for pointing this out. The training paradigm is indeed our core contribution. Prior methods that predict future states [2, 3, 4, 5, 6] train their world model with task-agnostic objectives or use it as a fixed simulator (Section 2), whereas SCIP (i) rewards plans by their imagined outcomes, (ii) trains the VLM to verify and revise its own plans, and (iii) calibrates the world model for plan verification. As suggested by the reviewer, we ablate each part on EB-Habitat. Retraining Section 3.3 without (i) or without (ii) lowers the average success rate by 4.0 and 6.0 (see the table of our response to (W2)), and (iii) adds 3.2 (64.8% to 68.0%, Table 4), indicating that each part of the training paradigm contributes to SCIP’s performance. We have added this analysis to Appendix C.1 of the revised paper.

[2] Cen et al. WorldVLA: Towards autoregressive action world model. arXiv 2025.

[3] Zhao et al. CoT-VLA: Visual chain-of-thought reasoning for vision-language-action models. CVPR 2025.

[4] Li et al. VLA-RFT: Vision-language-action reinforcement fine-tuning with verified rewards in world simulators. arXiv 2025.

[5] Xiao et al. World-Env: Leveraging world model as a virtual environment for VLA post-training. arXiv 2025.

[6] Wang et al. World modeling makes a better planner: Dual preference optimization for embodied task planning. ACL 2025.


Thank you for your valuable review and for taking the time to read our responses. All changes in the revised paper are marked in blue.

AuthorsOpenReview ↗

Official Comment by Authors

Thank you again for your time and consideration.

As the discussion period draws to a close, we would like to follow up on our response to your review. We would be grateful if you could let us know whether our response addresses your concerns. We would also be happy to provide further clarification or discuss any remaining questions.

Reviewer_YT9MOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SCIP, a framework in which a VLM agent self-corrects action plans before execution by checking them against imagined future visual states. It consists of three components: a model which generates policy, an imagination model that predicts the visual state change in latent space, and a verifier that judges whether the policy PASS or FAIL based on imaginary predictions. The framework is trained in three steps. First, supervised finetuning the VLM model with trajectories generated by advanced VLM models. Second step is using GRPO to train the policy model and verifier. The policy reward combines exact positional action match, cosine similarity between the final imagined latent of the candidate and ground-truth plans, and turn-wise improvement. Third, training the imagination model on frozen verifier to make verifier generate better judgment when it takes the prediction than without any prediction. On EmbodiedBench, SCIP reports the best average success rate among open-source embodied agents on all four environments.

Critical strengths

Training the world model so that its imagined states make plan verification more accurate is a clear, well-motivated idea. This is the paper’s main conceptual contribution and the main reason I see value in it. The plan-and-verify loop also looks like a useful idea. It helps even on 2B model without SFT (Table 7). Correcting failures before execution, which may be irreversible, is a relevant problem in embodied settings.

Critical weaknesses

The headline gains are hard to attribute to the proposed components. Most of the gain may come from SFT. The SFT-only model already reaches 57.6 on EB-Habitat (Tab. 4), well above every embodied baseline in Tab. 1. So, most of the reported +25.2% margin likely comes from SFT on EmbodiedBench trajectories distilled from 11 frontier VLMs, on the same environments and seen subsets, rather than from imagination or self- correction. The paper does not say whether the baselines (VAGEN, Vlaser, Reinforced Reasoner, LoHo-Manip, etc.) were trained or fine-tuned on the same EmbodiedBench data, re-run, or quoted from their papers. After Eq. 3, the ground-truth verdict is defined as y = PASS only if r_action = 1, which requires an exact action-by-action match with the single ground-truth plan. A valid plan that differs from the ground truth is therefore labeled FAIL. This contradicts the paper’s own motivation: r_visual is designed to credit any plan that reaches the correct final state, and Sec. 4.3 reports that most successful episodes use non-GT plans. These labels are then used to calibrate the world model (Sec. 3.4). Its objective rewards imagined states that help the frozen verifier predict y. It never compares them with the true next observation. The world model can thus improve verifier accuracy by encoding shortcut signals rather than accurate predictions; for valid non-GT plans, it is even pushed to imagine failure.

Questions for authors

The text following Eq. 3 defines y = PASS iff r_action = 1. Since r_action is a product of two factors bounded by 1, this holds only when an exact match with the single ground- truth plan. In contrast, the SFT verifier labels (App. D.3) appear to come from execution success. Please clarify which definition is used in Sec. 3.3 and 3.4, and how often valid non-GT plans are labeled FAIL under the RL definition. Given that y = PASS only for exact ground-truth matches, what is the verifier’s false- reject rate on plans that are valid but differ from the ground truth? How do the 68.5% / 50.4% non-GT successful plans get PASS? Were they executed after T=3 FAILs? Can you quantify the world model’s prediction error before vs. after verification-driven calibration?

What data was the world model pretrained on, and does it include scenes or episodes from the Common Sense / Spatial subsets? With 50 episodes per subset, the five-subset average on EB-Habitat must be a multiple of 0.4%. The T=1 value in Fig. 5 (64.5%) does not satisfy this, while the other three points do. Is it averaged over multiple runs, or is this a typo? Fig. 4 omits the Visual subset; from the reported averages it implies 82% Direct and 0% Recovery. Is there any reason why exclude Visual subset? How are Direct and Recovery defined at the episode level, given that an episode spans multiple planning steps? Fig. 4 attributes +6.8% to recovery, but Fig. 5 shows only +3.5% from T=1 to T=3, which suggests that roughly 3 points of episodes that succeed at T=1 fail at T=3, possibly due to false rejections by the verifier. Please report how many episodes are lost this way. Why is EmbodiedBench’s Long-Horizon subset excluded, given that the paper argues long-horizon instructions benefit most from imagined verification? The latency comparison is against Vlaser-8B, a different model. What is SCIP’s overhead relative to its own no-verification setting?

4 comments
AuthorsOpenReview ↗

Response to Reviewer YT9M (1/3)

(W1) How much of SCIP's gain comes from imagination and self-correction beyond SFT?

We appreciate the reviewer raising this concern. SFT only serves as SCIP’s starting point, and existing methods are also trained on data from these environments (e.g., Vlaser [1] on about 6M samples, including planning trajectories in the same Habitat simulator). On top of SFT and with the same data, imagination-based training (Sections 3.3 and 3.4) raises SCIP’s average success rate on EB-Habitat from 57.6% to 68.0% (+10.4), so this gain comes from the training, not the data.

The same training is also what puts SCIP ahead of Vlaser, the previous best open-source embodied agent, on the Common Sense and Spatial subsets, which are held out of VLM training. As shown in the table below, SFT alone barely exceeds Vlaser on Common Sense and falls below it on Spatial, while SCIP leads on both (+6.0 and +12.0 over SFT), indicating that most of SCIP’s margin over Vlaser on these subsets comes from imagination and self-correction, not SFT.

EB-Habitat SR (Tables 1 and 4)Common SenseSpatial
Vlaser (previous best, Table 1)32.026.0
Supervised fine-tuning (SFT)34.018.0
SCIP40.030.0
Gain over SFT+6.0+12.0

[1] Yang et al. Vlaser: Vision-language-action model with synergistic embodied reasoning. ICLR 2026.


(W2) Were the baselines (VAGEN, Vlaser, Reinforced Reasoner, LoHo-Manip, etc.) trained or fine-tuned on the same EmbodiedBench data, re-run, or quoted from their papers?

Thanks to the reviewer for requesting these details. The proprietary models are quoted from EmbodiedBench [2]. From ERA [3], we quote VAGEN [4] and Reinforced Reasoner [5] on EB-ALFRED, RL4VLM [6], and Robot-R1 [7]. From their own papers, we quote Reinforced Reasoner on EB-Habitat and LoHo-Manip [8]. We run the rest ourselves. Vlaser [1] is evaluated from its released 8B checkpoint, with its EB-ALFRED average taken from its paper, while VAGEN (EB-Habitat, EB-Navigation) and VLA-RFT [9] are trained with their official code. As for training data, VAGEN, RL4VLM, and VLA-RFT are trained in the environment they are evaluated on, while the other baselines are used as trained in their own papers. SCIP outperforms every baseline trained in the evaluated environment (Tables 1 to 3), and Table 4 compares SCIP with its own SFT on identical data. We have stated the source of each baseline in Section 4.1 of the revised paper.

[2] Yang et al. EmbodiedBench: Comprehensive benchmarking multi-modal large language models for vision-driven embodied agents. ICML 2025.

[3] Chen et al. ERA: Transforming VLMs into embodied agents via embodied prior learning and online reinforcement learning. arXiv 2025.

[4] Wang et al. VAGEN: Reinforcing world model reasoning for multi-turn VLM agents. NeurIPS 2025.

[5] Wu et al. Reinforced reasoning for embodied planning. arXiv 2025.

[6] Zhai et al. Fine-tuning large vision-language models as decision-making agents via reinforcement learning. NeurIPS 2024.

[7] Kim et al. Robot-R1: Reinforcement learning for enhanced embodied reasoning in robotics. NeurIPS 2025.

[8] Liu, Isabella, et al. "Long-horizon manipulation via trace-conditioned vla planning." CoRL 2026.

[9] Li, Hengtao, et al. VLA-RFT: Vision-language-action reinforcement fine-tuning with verified rewards in world simulators. arXiv 2025.


(W3, Q1) The verdict label is PASS only for an exact match with the ground-truth plan, whereas r_visual credits any plan reaching the correct final state and SFT labels use execution success (Appendix D.3). Which label do Sections 3.3 and 3.4 use, and is it consistent with r_visual?

Thanks to the reviewer for the opportunity to clarify this. Sections 3.3 and 3.4 both use the exact-match label. SFT can use execution success because its plans come from recorded trajectories (Appendix D.3), whereas these two stages train on plans generated on the fly, and labeling them by execution would require running the simulator on every sampled plan, which is time-consuming. Only the ground-truth plan is therefore labeled PASS, and any other plan FAIL.

This label does not conflict with the visual reward, because the reward trains the planner and the label trains the verifier. It also does not make the verifier require an exact match. The verifier takes the imagined outcome of a plan as input, and the ground-truth plan only defines its training target, so it learns to judge a plan by this outcome. In the EB-Habitat evaluation, the verifier accepts 180 of the 186 valid plans that differ from the ground-truth plan (96.8%, see our response to (Q2)), indicating that it does not require an exact match to accept a plan.

AuthorsOpenReview ↗

Response to Reviewer YT9M (2/3)

(Q2) How often does the verifier reject valid plans that differ from the ground-truth plan?

We thank the reviewer for the question. To measure this, we replayed the rejected plans of the EB-Habitat evaluation from the state where they were proposed. This state can be reproduced exactly for 75 of the 114 rejected plans, which are proposed after the first action of an episode, as the initial pose of the robot is random. Only 6 of these 75 plans are valid, while the other 69 (92%) would have failed. Among the plans proposed after the first action, 186 valid ones differ from the ground-truth plan, and these 6 are the only ones the verifier rejects (3.2%), indicating that it rarely rejects a valid plan.


(Q3) (1) How do the successful plans that differ from the ground-truth plan (68.5% / 50.4%) get PASS? (2) Were they executed after three FAILs?

We thank the reviewer for the opportunity to clarify this. (1) The verifier passes them itself, because it judges a plan by its imagined outcome, not by a match with the ground-truth plan, as explained in our response to (W3, Q1). (2) No, they are executed after a PASS from the verifier. Plans executed after three FAILs account for only 1.2% of the executed plans on EB-Habitat, and in that case the planner revises the plan once more with the last feedback before executing it (Section 3.5).


(W4, Q4) What is the world model's prediction error before and after calibration (Section 3.4)? Does calibration keep imagined states close to the real observations, or push valid non-GT plans toward imagined failure?

Thanks to the reviewer for raising this. We quantify the prediction error by the L1 error between imagined states and real observations in the feature space of the world model (V-JEPA 2 [10]), on the plans that SCIP executes in the EB-Habitat evaluation. As shown in the table below, calibration with verification signals alone raises this error from 0.351 to 0.850 at 4 steps, even though the success rate rises from 64.8% to 68.0% (Table 4). A higher success rate therefore does not ensure that imagined states stay close to the real observations. To ensure this, we add a next-observation loss to calibration (Section 3.4), which trains the world model to match the real next observation on the training trajectories. With this loss, the error returns to the level before calibration (0.358 vs. 0.351), while most of the gain in success rate is retained (67.7% vs. 68.0%), indicating that the revised calibration keeps imagined states close to the real observations instead of relying on shortcut signals. We have added this loss to Section 3.4 and this analysis to Section 4.3 (Table 6) of the revised paper.

Pre-training already minimizes this error (Appendix D.4), so calibration serves a different goal, which is to make imagined states support correct plan verification (Section 2). As calibration updates only the world model with the VLM fixed (Section 3.5), its gain comes entirely from the imagined states. With the same VLM, the average success rate is 63.6% without imagined states (Table 5), 64.8% with the imagined states before calibration (Table 4), and 67.7% after the revised calibration, at the same error level, indicating that calibration makes imagined states more useful for verification without moving them away from the real observations.

This also holds for valid non-GT plans. Among the executed plans, those that differ from the ground-truth plan are imagined no less accurately than those that match it (0.343 vs. 0.386 at 4 steps). Among the 186 valid plans that differ from the ground-truth plan, the verifier rejects only 6 (see our response to (Q2)), indicating that the world model rarely makes a valid plan look like a failure.

EB-HabitatL1 ↓, 1 step4 stepsAvg. SR ↑
Before calibration0.2240.351–
SCIP (after calibration)0.3510.85068.0
SCIP + next-observation loss0.2250.35867.7

SR averaged over three runs.

[10] Assran et al. V-JEPA 2: Self-supervised video models enable understanding, prediction and planning. arXiv 2025.

AuthorsOpenReview ↗

Response to Reviewer YT9M (3/3)

(Q5) (1) What data was the world model pretrained on, and (2) does it include scenes or episodes from the Common Sense / Spatial subsets?

We thank the reviewer for the question. (1) The world model is pretrained on (observation, action, next observation) transitions from the frontier-model trajectories released with EmbodiedBench [2], without task instructions or success labels (Appendices D.3 and D.4). (2) These transitions come from all five subsets. Nevertheless, this does not expose the Common Sense and Spatial tasks, because the world model takes only an observation and an action as input, so it learns how actions change a scene, not how a task is solved. The planner, which reads the instruction, is trained only on the Base, Complex, and Visual subsets (Appendix D.3). We have clarified this in Section 4.1 and Appendix D.4 of the revised paper.


(Q6) The T = 1 value in Fig. 5 (64.5%) is not a multiple of 0.4%, as a five-subset average over 50 episodes each must be. Is it averaged over multiple runs, or is this a typo?

We thank the reviewer for catching this typo. We re-ran SCIP three times each with T = 1, 2, and 3 on all 250 EB-Habitat episodes, reaching 64.8%, 66.8%, and 68.0% on average, which indicates that the trend of Fig. 5 holds. We have updated Fig. 5 and the corresponding text in Section 4.3 with these three-run averages.


(Q7) Fig. 4 omits the Visual subset, which implies 82% Direct and 0% Recovery. Why is it excluded?

We thank the reviewer for the question. We omitted Visual because all of SCIP’s successes there are Direct (82.0%), with none after revision. This indicates that SCIP’s gain over SFT on Visual (82.0% vs. 60.0%, Table 4) comes entirely from better first plans, which also give SCIP the best EB-Habitat Visual score in Table 1.


(Q8) How are Direct and Recovery defined at the episode level, given that an episode spans multiple planning steps?

Thanks to the reviewer for requesting these details. An episode is labeled over all of its planning steps. It is Direct if it succeeds without any revised plan, Recovery if it succeeds with at least one plan revised after a rejection, and Failure otherwise, so every episode falls into exactly one category. All numbers in Fig. 4 follow these definitions.


(Q9) Fig. 4 attributes +6.8% to recovery, but Fig. 5 shows only +3.5% from T = 1 to T = 3. How many episodes that succeed with T = 1 fail with T = 3 due to false rejections?

We thank the reviewer for the question. Only 5 of the 250 EB-Habitat episodes (2.0%) succeed with T = 1 but fail with T = 3, while 10 episodes fail with T = 1 but succeed with T = 3. This is a net gain of 5 episodes, indicating that the extra verification turns help more than they hurt. Here, an episode counts as a success under a setting if it succeeds in at least two of our three runs of that setting. Note that the Recovery in Fig. 4 (6.8%) is not expected to equal the gain from T = 1 to T = 3 in Fig. 5 (+3.2% after the correction in (Q6)), because with T = 1 a rejected plan is still revised once, so part of this Recovery is already obtained at T = 1.


(Q10) Why is EmbodiedBench's Long-Horizon subset excluded, given that long-horizon instructions benefit most from imagined verification?

Thanks to the reviewer for raising this. We excluded Long-Horizon because we follow the evaluation protocol and baselines of ERA [3], which evaluates the five other subsets. As suggested by the reviewer, we additionally evaluated SCIP on the Long-Horizon subset of EB-Habitat, which its planner never sees in training. As shown in the table below, SCIP reaches 30.0%, while LoHo-Manip-4B [8], an embodied agent built for long-horizon tasks, reaches 20.0% as reported in its paper, indicating that SCIP remains effective on long-horizon tasks.

MethodLong-Horizon SR
LoHo-Manip-4B (reported in its paper)20.0
SCIP30.0

(Q11) The latency comparison is against Vlaser-8B, a different model. What is SCIP's overhead relative to its own no-verification setting?

We thank the reviewer for the question. Compared with its own no-verification setting, SCIP adds 0.60 s per planning step on average, less than half of the 1.36 s that generating a plan takes. We measured this on EB-Habitat with one H200 GPU.


Thank you for your valuable review and for taking the time to read our responses. All changes in the revised paper are marked in blue.

AuthorsOpenReview ↗

Official Comment by Authors

Thank you again for your time and consideration.

As the discussion period draws to a close, we would like to follow up on our response to your review. We would be grateful if you could let us know whether our response addresses your concerns. We would also be happy to provide further clarification or discuss any remaining questions.

Reviewer_e1N5OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces proposes a framework for training embodied agents, called SCIP. In this framework the model verify and revise plans before execution using imagined future visual states. In particular, SCIP uses a learned world model to predict the visual consequences of actions and a verifier to determine whether the plan is likely to succeed. If the plan is judged invalid, the agent refines it before acting. This self-corrective process is used at both training and inference phases. The model is tested on EmbodiedBench tasks including navigation and manipulation and show improvements in the related performance metrics.

Critical strengths

1- The paper has a novel and well-motivated integration of imagination-based world-modeling and self-correction for embodied agents. 2- It shows consistent improvement on various tasks for navigation and object manipulation. 3- Ablates the contribution of each components and shows both imagination and verification steps contribute to the improvements in the results.

Critical weaknesses

-Evaluation is limited to the EmbodiedBench. -It is unclear whether the learned imagination model captures transferable action dynamics or environment-specific patterns. -quantitative analysis of world-model accuracy and long-horizon rollout errors is limited. -Limited discussion of failure mode

Questions for authors

-Since SCIP is trained using task-specific trajectories and environment interactions, to what extent does the learned imagination model capture general action-environment dynamics versus overfitting to the visual characteristics and transition patterns of the training environments? In particular, is the improvement primarily due to better modeling of environment-specific visual features, or does the approach lead to more transferable reasoning about action consequences and state transitions?

-Since the proposed approach relies heavily on imagined future states, it would be valuable to quantify prediction quality more directly. The paper provides limited analysis of long-horizon rollout errors, error accumulation, or the relationship between imagination accuracy and downstream task success. Can you discuss this further?

-It would be helpful to see the computational cost/overhead of each step that is imagination and verification. How many rounds of verification is needed and is this scalable for inference time? Can we also isolate the impact of the approach when it is used at training only or at inference only?

4 comments
AuthorsOpenReview ↗

Response to Reviewer e1N5 (1/3)

(W1) Evaluation is limited to EmbodiedBench.

Thanks to the reviewer for raising this. We evaluate SCIP on EmbodiedBench [1] because recent embodied agents are also compared on it (LoHo-Manip [2], Vlaser [3], VAGEN [4], Robot-R1 [5], RL4VLM [6], etc.), which allows a direct comparison. Further, it is diverse, spanning four environments (two household-task simulators, navigation, and robot-arm control), across high-level planning and low-level control tasks.

To further widen the evaluation as suggested by the reviewer, we additionally evaluated SCIP on the Long-Horizon subset of EB-Habitat [1], which the paper does not cover. As shown in the table below, SCIP reaches 30.0% with a 2B backbone, while LoHo-Manip-4B [2], an embodied agent built for long-horizon tasks, reaches 20.0% as reported in its paper, indicating that SCIP remains effective on long-horizon tasks. Beyond this subset, SCIP also achieves the best average success rate among open-source embodied agents in all four EmbodiedBench environments (Tables 1 to 3).

MethodLong-Horizon SR
LoHo-Manip-4B [2]20.0
SCIP30.0

[1] Yang, Rui, et al. EmbodiedBench: Comprehensive benchmarking multi-modal large language models for vision-driven embodied agents. ICML 2025.

[2] Liu, Isabella, et al. "Long-horizon manipulation via trace-conditioned vla planning." CoRL 2026.

[3] Yang, Ganlin, et al. Vlaser: Vision-language-action model with synergistic embodied reasoning. ICLR 2026.

[4] Wang, Kangrui, et al. VAGEN: Reinforcing world model reasoning for multi-turn VLM agents. NeurIPS 2025.

[5] Kim, Dongyoung, et al. Robot-R1: Reinforcement learning for enhanced embodied reasoning in robotics. NeurIPS 2025.

[6] Zhai, Yuexiang, et al. Fine-tuning large vision-language models as decision-making agents via reinforcement learning. NeurIPS 2024.

[7] Bai, Shuai, et al. Qwen3-VL technical report. arXiv 2025.


(W2, Q1) (1) Does the learned imagination model capture transferable action dynamics or environment-specific patterns? (2) Is the improvement primarily due to better modeling of environment-specific visual features, or to more transferable reasoning about action consequences and state transitions?

Thanks to the reviewer for the opportunity to clarify this.

(1) Our imagination model (the world model) captures both transferable action dynamics and environment-specific patterns. We first show that it captures transferable action dynamics. Its input is only an observation and an action, without the task instruction (Section 3.2). It therefore learns how an action changes the scene, not how a task is solved, and this does not depend on the specific planner. As shown in the table below, when we plug it into the original Qwen3-VL-2B [7] without any fine-tuning, the success rate rises from 11.2% to 20.8% on EB-Habitat and from 4.8% to 10.4% on EB-ALFRED, indicating that the action dynamics captured by the world model can be utilized by a planner never trained with it.

We now clarify how our method also captures environment-specific patterns, as the imagination model is pre-trained on the transitions of each environment (Appendix D.4). Qualitatively, its decoded imagined states match the real observations (Fig. 7) and follow the rules of each environment (e.g., putting down an object while holding none leaves the imagined state unchanged, in Fig. 8). Therefore, these two are complementary: the action dynamics tell the world model what an action changes, and the environment-specific patterns tell it how the result looks in that environment.

Qwen3-VL-2B, no fine-tuning (Table 8)EB-Habitat SREB-ALFRED SR
Without world model11.24.8
With world model20.810.4

(2) The improvement mainly comes from more transferable reasoning about action consequences. In SCIP, modeling visual features is the role of the world model, and reasoning about action consequences is the role of the VLM, which plans and verifies with the imagined consequences. As shown in the table below, most of the improvement over supervised fine-tuning (SFT) appears in policy learning (+7.2 of +10.4 on average), which trains only the VLM while keeping the world model fixed. The improvement also reaches the Common Sense and Spatial subsets (+6.0 and +12.0 over SFT), whose instructions the VLM never sees in training, indicating that this improvement mainly comes from reasoning of the VLM that transfers to unseen instructions, instead of better modeling of environment-specific visual features.

EB-Habitat SR (Table 4)Avg.Common SenseSpatial
Supervised fine-tuning57.634.018.0
+ Policy learning (Sec. 3.3)64.8 (+7.2)40.026.0
+ Calibration (Sec. 3.4)68.0 (+3.2)40.030.0
AuthorsOpenReview ↗

Response to Reviewer e1N5 (2/3)

(W3, Q2) Can you quantify the accuracy of the world model more directly, including long-horizon rollout errors, error accumulation, and the relation between imagination accuracy and task success?

Thanks to the reviewer for raising this. We quantify the accuracy of the world model by the L1 error between its imagined states and the real observations in its feature space (V-JEPA 2 [8]). We measure it by rolling out 1 to 4 consecutive actions of each plan that SCIP executes in the EB-Habitat evaluation. As shown in the table below, calibration with verification signals alone raises this error and lets it accumulate over steps (from 0.351 at 1 step to 0.850 at 4 steps). We therefore add a next-observation loss to the calibration of the world model (Section 3.4), which trains it to match the real next observation on the training trajectories. With the next-observation loss, the error returns to the level before calibration (0.225 vs. 0.224 at 1 step and 0.358 vs. 0.351 at 4 steps) and stays flat from 2 to 4 steps (0.355, 0.348, and 0.358), indicating that the imagined states still correspond to the actual environment dynamics and the error does not accumulate over longer rollouts.

As for the relation to task success, imagined states contribute to it, since removing them from the verifier lowers the average success rate on EB-Habitat from 68.0% to 63.6% (Table 5). Keeping them close to real observations preserves this benefit, as calibration with the next-observation loss keeps its gain in success rate (67.7% vs. 68.0% without this loss). We have added this loss to Section 3.4 and this analysis to Section 4.3 (Table 6) of the revised paper.

L1 error ↓1 step2 steps3 steps4 stepsAvg. SR
Before calibration0.2240.3550.3450.351-
SCIP (after calibration)0.3510.5680.7060.85068.0
SCIP + next-observation loss0.2250.3550.3480.35867.7

SR averaged over three runs.

[8] Assran et al. V-JEPA 2: Self-supervised video models enable understanding, prediction and planning. arXiv 2025.

AuthorsOpenReview ↗

Response to Reviewer e1N5 (3/3)

(W4) Discussion of failure modes.

Thanks to the reviewer for pointing this out. As suggested, we additionally analyze the failure modes of SCIP on EB-Habitat by checking every failed episode against its goal. As shown in the table below, 65 of the 80 failures across all five subsets (81%) come from the Common Sense and Spatial subsets, and 59 of these 65 differ from the goal only in the object (29 of 30 in Common Sense) or the place (30 of 35 in Spatial). For example, the robot brings a toy airplane for “something to play with”, where the intended target is the ball. The rest of the plan is correct, and what fails is identifying a target that the instruction names only indirectly.

Failed episodesCount
All subsets80
Common Sense + Spatial65 (81%)
Common Sense, different object only29 of 30
Spatial, different place only30 of 35

Identifying an indirectly named target is hard even for strong proprietary models, which are not trained on any EmbodiedBench data. On EB-Habitat, GPT-4o falls from 86.0% on Base to 44.0% on Common Sense and 36.0% on Spatial, and Claude-3.7-Sonnet from 90.0% to 58.0% and 38.0% (Table 1), indicating that this difficulty is common across models. On these two subsets, SCIP still reaches 40.0% and 30.0% with a 2B backbone, the highest among open-source embodied agents (Table 1). We have added this analysis to Appendix B.3 of the revised paper.


(Q3) (1) What is the computational cost of imagination and verification? (2) How many rounds of verification are needed, and is this scalable at inference time? (3) Can the impact of the approach be isolated when it is used at training only or at inference only?

We thank the reviewer for these questions.

(1) Imagination and verification together cost less per planning step than generating the plan. As shown in the table below, they take 0.16 s and 0.27 s, against 1.36 s for plan generation. The difference comes from the output length. The planner generates a state description, its detailed reasoning, and a plan composed of multiple actions, while the verifier only needs to reason over the imagined states and generates a verdict with a concise explanation.

Inference costValue
Plan generation, per step1.36 s
Imagination, per step0.16 s
Verification, per step0.27 s

(2) Three turns are sufficient. The average SR over three runs is 64.8%, 66.8%, and 68.0% with T = 1, 2, and 3 on EB-Habitat, and a fourth turn brings no further gain (Fig. 5). The loop is also scalable at inference. Since the number of turns is capped at three and a further turn runs only when the verifier rejects a plan, the whole loop, including the revision of rejected plans, adds 0.60 s per planning step on average, less than half of the 1.36 s for generating a plan.

(3) Yes. Used at inference only, imagination raises the success rate of the original Qwen3-VL-2B [7], which was never trained with it, from 11.2% to 20.8% on EB-Habitat and from 4.8% to 10.4% on EB-ALFRED (Table 8). Used in training only, with the imagined states removed from the verifier at inference, SCIP reaches 63.6% on EB-Habitat, above SFT (57.6%, Table 4) while below the full SCIP (68.0%, Table 5). These indicate that imagination helps across training and inference stages and helps most when used at both.


Thank you for your valuable review and for taking the time to read our responses. All changes in the revised paper are marked in blue.

AuthorsOpenReview ↗

Official Comment by Authors

Thank you again for your time and consideration.

As the discussion period draws to a close, we would like to follow up on our response to your review. We would be grateful if you could let us know whether our response addresses your concerns. We would also be happy to provide further clarification or discuss any remaining questions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely thank the AC and reviewers for their valuable time and insightful feedback, which substantially strengthened our work. The strengths pointed out include:

  1. (Motivation): A well-formulated problem relevant to embodied settings. Correcting failures before execution, which may be irreversible, is a relevant problem in embodied settings [YT9M], and the problem is well formulated [6cFd].

  2. (Method): A novel, well-motivated, and intuitive framework. SCIP is a novel and well-motivated integration of imagination-based world modeling and self-correction for embodied agents [e1N5]. Training the world model so that its imagined states make plan verification more accurate is a clear and well-motivated idea, and the paper's main conceptual contribution [YT9M]. The integration of planning, latent prediction, verification, and revision is intuitive and clearly motivated, and the framework is trained with a novel paradigm [6cFd].

  3. (Experiment/Performance): Consistent and broad improvement, even without SFT. SCIP achieves consistent improvement on various navigation and manipulation tasks, and the ablation shows that both imagination and verification contribute [e1N5]. The plan-and-verify loop looks like a useful idea and helps even a 2B model without SFT [YT9M]. The experiments cover both high-level planning and low-level control [6cFd].

  4. (Presentation): Clear presentation. The paper is well-written [6cFd], clear, and easy to follow [e1N5, 6cFd].

The concerns raised by the reviewers are summarized below. Please refer to our individual responses for further details.

  1. Analysis of whether the world model is faithful to the actual environment dynamics and transferable [W2-W3, Q1-Q2 by e1N5; W4, Q4 by YT9M; W3 by 6cFd]

  2. Contribution of imagination-guided learning beyond SFT, the two-stage schedule, and distinction from prior methods [W1 by YT9M; W1, W2, W4 by 6cFd]

  3. Discussion of evaluation scope, the Long-Horizon subset, and failure modes [W1, W4 by e1N5; Q10 by YT9M]

  4. Analysis of the cost and effect of imagination and verification [Q3 by e1N5; Q11 by YT9M]

  5. Clarification of the verdict label for valid non-GT plans, baseline sources, pre-training data, and the numbers in Figs. 4-5 [W2-W3, Q1-Q3, Q5-Q9 by YT9M]

Our rebuttal adds (i) a next-observation loss with a quantitative analysis of the faithfulness of imagined states (Table 6), (ii) an ablation of the visual-similarity reward and verifier training in Section 3.3, (iii) an evaluation on the Long-Horizon subset, and (iv) an analysis of failure modes (Appendix B.3). Building on the strengths recognized by all reviewers and these improvements, we believe this work offers significant value to the embodied AI community, particularly for self-correcting planning with world modeling, and respectfully hope they will be considered in the final evaluation. Thank you again for your kind attention and assistance.

Sincerely,

Authors of Submission 9268

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.