acceptodds
Under review as a conference paper at ICLR 2027

Think Fast, Plan Selectively: Adaptive Deliberation for Efficient Data-Driven MPC

Abstract

Data-driven model predictive control (MPC) combines learned world models with online trajectory optimization, achieving strong performance in continuous control. However, the per-step cost of sampling and evaluating hundreds of candidate trajectories restricts deployment to control frequencies well below what real-time robotics demands. Motivated by the dual-process theory of human cognition, which distinguishes between fast, intuitive processing (System 1) and slower, deliberative reasoning (System 2), we ask whether every decision requires the same degree of computational deliberation. We propose Fast-TD-MPC, a lightweight framework that adaptively routes between fast policy execution and test-time planning, reserving costly deliberation for states where it is most needed. Fast-TD-MPC delivers competitive task performance across 103 continuous control tasks while achieving up to  4x faster inference. Under external disturbances, Fast-TD-MPC selectively falls back to planning, maintaining robustness comparable to the original planner.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Model-Based Planning, Adaptive Computation, Test-Time Planning, TD-MPC

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_wj9NOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Fast-TD-MPC, a fast–slow architecture for efficient test-time planning. It combines a lightweight neural policy (System 1), trained to imitate the TD-MPC2 planner, with the original MPPI planner (System 2). A latent-space OOD detector determines when to invoke the expensive planner instead of the fast policy. The authors evaluate the approach across 103 simulated continuous-control tasks, demonstrating improvements in inference efficiency while maintaining competitive task performance.

Critical strengths
  • Promising Systems Idea: With the increasing use of world models and reasoning models, fast–slow architectures represent a promising direction for improving computational efficiency. I appreciate the authors' efforts to explore such an architecture for efficient test-time planning and control.
  • Extensive Simulation Experiments, Comparisons, and Ablation Studies: I appreciate the authors' efforts in conducting extensive simulation experiments across diverse tasks, along with several baseline comparisons and ablation studies.
Critical weaknesses
  • Novelty: My primary concern with the paper is its methodological novelty. While the overall idea and system architecture are promising, I find it difficult to identify the specific novel contribution. The current presentation appears to emphasize the System 1–System 2 architecture as a key contribution. However, as the authors themselves acknowledge, prior works such as When to Plan and AdaReP have already explored related approaches to selective planning. Although the specific implementations differ, the novelty of the proposed framework relative to these existing methods remains unclear. I would encourage the authors to clarify and better articulate the novel contributions of this work.
  • Real-World Experiments: One of the primary advantages of reducing computational complexity is its potential impact on real-time robotic applications, particularly when deployed on resource-constrained edge devices. However, the current evaluation is limited to simulation, where the practical impact of real-time computational constraints is less evident. Do the authors have any real-world experiment results to demonstrate the effectiveness of the approach
  • OOD Detector Design: If my understanding is correct, the proposed OOD detector characterizes the latent-state distribution for each task using a single mean and covariance matrix, effectively approximating it with a single Gaussian representation. This seems to be an oversimplification, as the latent-state distribution of a task may be multimodal. Could the authors clarify the reasoning behind this design choice and justify whether such a representation is sufficient for the considered tasks?
Questions for authors

The weakness mentions the questions

3 comments
AuthorsOpenReview ↗

Response to Question Item 1-3

Thank you for the comments.

Q1: Methodological novelty beyond existing selective-planning methods is unclear

Our contribution is how actions are produced when planning is skipped: S1 computes every such action directly from the current observation, whereas AdaReP reuses cached plans [1] and Lin et al. correct queued planner actions [2]. Our gate is fitted offline from latent statistics, rather than trained through RL as in When to Plan [3].

The method preserves the pretrained TD-MPC2 world model and planner, with one training configuration across 103 tasks. The matched-latency controls show why the routing matters: on DMControl, we achieve 785 vs. 743 return at 6.5/6.6 ms against periodic planning. On disturbed DMControl, the same threshold increases planning from 18% to 45% while retaining 95% of planner return at half its mean latency.

Q2: Are there real-world results demonstrating benefits for real-time robotic control?

We’d like to clarify that our contribution is a fast policy that operates without a cached plan. S1 learns the planner’s executed actions and generates an action from each current observation. AdaReP instead reuses cached plans [1], while Lin et al. correct queued actions using observation feedback [2]. Our gate is fitted offline from latent statistics, without the RL-trained routing policy used in When to Plan [3].

The method preserves the pretrained TD-MPC2 world model and planner, using one training configuration across 103 tasks. On DMControl, the same threshold increases planning from 18% to 45% under disturbances while retaining 95% of planner return at half its mean step latency.

We'll expand Appx. L with a comparison of the fast action generator, routing signal and training requirements.

Q3: Why use one mean and covariance, and is this sufficient for tasks with potentially multimodal latent distributions?

We use the mean and full covariance as a low-cost routing statistic, fitted per task-seed. The covariance accounts for correlations and different scales of latent variation. Thresholds are calibrated from held-out empirical scores, so the routing rule does not require Gaussian-distributed latents. Feature-space OOD and anomaly detection provide precedents for this approach [4,5].

A task can still have multiple modes. Our aim is to obtain a useful routing signal, rather than reconstruct the complete latent density. We assess this through control performance and computation. On DMControl, the gate achieves 785 return at 6.5 ms versus 743 at 6.6 ms for periodic planning, with a scoring cost of 0.17 ms on H100.

Appx. J also compares alternative distance metrics and shows that the choice affects return and planning frequency. These results support the chosen statistic for the reported performance–latency trade-offs.

References:

[1] Cheng et al. AdaReP: Adaptive Re-Planning under Model Mismatch for Neural World-Model Predictive Control. ICANN'26.

[2] Lin et al. Accelerating Multi-modal LLM Gaming Performance via Input Prediction and Mishit Correction. arXiv'25.

[3] Labiosa and Hanna. When to Plan: Learning to Select Between Reactive Control and Deliberative Planning. RLC'26.

[4] Lee et al. A Simple Unified Framework for Detecting Out-of-Distribution Samples and Adversarial Attacks. NeurIPS'18.

[5] Rippel et al. Modeling the Distribution of Normal Data in Pre-Trained Deep Features for Anomaly Detection. ICPR'20.

Thank you for the rebuttal. Most of my conceptual comments were addressed but the real world demonstration is needed to understand if the impact found in simulation is transferable.

Dear Reviewer wj9N,

Thank you for your insightful comments.

Sincerely,
The authors

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

This paper presents Fast-TD-MPC, a framework for reducing the inference cost of a pretrained TD-MPC2 controller through selective planning. A compact policy is trained to imitate the planner, while a Mahalanobis-distance gate in the frozen latent space routes familiar states to this policy and atypical states to the original MPPI planner. The gate is calibrated offline using either all policy-generated states or a reward-filtered subset. Evaluation covers 103 continuous-control tasks across four domains with three seeds, together with combined-disturbance experiments on 102 tasks. The strongest results are on DMControl, where the method achieves approximately 4.2× lower average step latency while retaining 96% of the planner’s return, and approximately 2× speedup while retaining 95% under disturbances. Benefits vary across domains, and fixed-frequency planning remains competitive in some settings.

Critical strengths

The main strengths are the practical simplicity of the framework and the breadth of its empirical evaluation. Together, they make selective planning a useful and reasonably well-supported inference strategy.

  1. The framework addresses a concrete inefficiency in learned MPC: repeatedly performing expensive trajectory optimization when a reactive policy can already produce adequate actions. It preserves the pretrained world model and planner, adds a compact distilled policy and inexpensive gate, and retains observation feedback at every step. This makes the approach straightforward to integrate and provides a transparent mechanism for trading computation against control performance.

  2. The evaluation includes 103 tasks, three seeds, threshold sweeps, component-level timing, fixed-frequency planning baselines, and disturbance tests. On DMControl, the method obtains 785 return at 6.5 ms versus 743 at 6.6 ms for round-robin planning, providing evidence beyond simply reducing planning frequency. Under disturbances, the same configuration increases planner usage and retains 95% of the full planner’s return at approximately half the latency. These results substantiate a practically meaningful benefit in at least one substantial task domain.

Critical weaknesses

The main reservations concern the incremental contribution over existing adaptive planning methods and the scope of the real-time claims. These limit the strength of the recommendation, despite the useful empirical results. Additionally, a lack of real world deployment is also a critical perspective of weakness.

  1. The method combines established planner distillation and distance-based OOD detection, but the evaluation omits the closest adaptive replanning approaches discussed in the paper, including AdaReP and speculation-and-correction methods. Comparisons with fixed scheduling are informative but do not fully establish the value of this particular routing design. Moreover, round-robin planning matches or outperforms gating in parts of the evaluation, especially MyoSuite. The evidence therefore supports a useful implementation more strongly than a broadly superior adaptive planning mechanism.

  2. The real-time motivation is only partially validated. Reported latency is averaged over fast and slow steps and includes environment execution, while each fallback still incurs the cost of full MPPI planning. The experiments do not establish tail latency, deadline compliance, or control performance when the environment continues evolving during computation. Consequently, the demonstrated reduction in average computation does not yet establish mitigation of the stale-action problem emphasized in the introduction.

  3. All experiments are conducted in simulation, leaving the practical benefits for real-time robotic control unverified on physical systems. In particular, it remains unclear whether the reactive OOD gate can trigger planning early enough to maintain reliable control under real sensor noise, actuation delays, and model mismatch.

Questions for authors
  1. How does Fast-TD-MPC compare with AdaReP or a mismatch-based speculation-and-correction baseline under the same world model, hardware, and matched latency or planning budget? Does the advantage persist beyond fixed-frequency scheduling?

  2. How strongly does the OOD score predict the additional control benefit of invoking S2 instead of S1? In particular, the “all” mode includes unsuccessful S1 states in its reference distribution; can paired rollout analysis show that high-scoring states are actually those where planning is more beneficial?

  3. What are the controller-only median and tail latencies, and what fraction of steps meet a specified control deadline? Do the observed benefits persist when computation-induced action delays affect environment evolution, and what execution optimizations are enabled for the MPPI baseline?

4 comments
AuthorsOpenReview ↗

Response for Question Item 1-4

Thank you for the thoughtful questions.

Q1: What distinguishes Fast-TD-MPC, and how does it compare with adaptive baselines under matched conditions?

We’d like to clarify that our fast policy operates independently of a cached plan; whereas AdaReP adapts cached-plan reuse [1], while Lin et al. correct queued actions using observation feedback [2]. S1 instead generates actions from the current observation through planner imitation, so its execution does not depend on a queue’s remaining horizon. Our gate uses offline latent statistics without an RL-trained routing policy [3], while retaining the full pretrained planner for fallback.

Our evaluation tests this design across 103 tasks. The fixed-frequency controls hold the world model, planner and hardware constant, testing whether state-dependent allocation improves the performance–compute trade-off. We’ll expand Appx. L with a side-by-side comparison of the fast action generator, routing signal and training requirements.

Q2: Does the OOD score predict the additional benefit of invoking S2?

We use the OOD score as a low-cost proxy for S1 reliability. This follows confidence-based expert intervention in EnsembleDAgger [5] and uncertainty-guided selection between reactive control and planning in When to Plan [3].

We evaluate this proxy through the resulting performance–compute trade-off. On DMControl, routing achieves 785 return with 18% planning, versus 743 with 20% planning for RR at similar latency. Under disturbances, the same threshold increases planning to 45% and retains 95% of planner return at half its mean step latency.

The all-mode reference can include unsuccessful S1 states. Our reward-gated mode adds performance information by filtering the reference using expert-relative per-step rewards or completion proxies. The score serves as an intervention signal, rather than a numerical estimate of the return gained from each planning call.

Q3: What are the controller-only latencies, deadline compliance, delay effects and MPPI optimizations?

We’d like to clarify that our experiments evaluate the performance–computation trade-off using synchronized mean step latency. Table 2 separates the component costs, and the synchronous protocol provides consistent timing conditions across methods.

Fallback steps retain full MPPI cost, so the demonstrated benefit is lower average execution cost rather than a guaranteed increase in control frequency. We will clarify this distinction in the real-time motivation. MPPI uses the same uncompiled implementation across the planning-based comparisons.

Q4: Physical deployment remains unverified

We’d like to clarify that this study provides a proof of concept in simulation. Controlled comparisons across 103 tasks, with disturbance tests on 102 tasks, evaluate the performance–computation trade-off of selective planning.

Prior physical deployments of TD-MPC-based controllers [4] motivate extending Fast-TD-MPC to physical systems. We leave this evaluation to future work, including robustness to real sensor noise, actuation delays, model mismatch and the timing of gate interventions.

References:

[1] Cheng et al. AdaReP: Adaptive Re-Planning under Model Mismatch for Neural World-Model Predictive Control. ICANN'26.

[2] Lin et al. Accelerating Multi-modal LLM Gaming Performance via Input Prediction and Mishit Correction. arXiv'25.

[3] Labiosa and Hanna. When to Plan: Learning to Select Between Reactive Control and Deliberative Planning. RLC'26.

[4] Feng et al. Finetuning Offline World Models in the Real World. CoRL'23.

[5] Menda et al. EnsembleDAgger: A Bayesian Approach to Safe Imitation Learning. IROS’19.

Thank you for the response. The response has addressed some of my concerns, but still I think real world deployment is critical for some of the manipulation tasks, I will keep my score as 3.

By the way I couldn’t see Fig.1 in the pdf file, there might be some rendering issues and the authors might need to check it.

Dear Reviewer Fbsj,

Thank you for your valuable comments.

For Figure 1, we've checked the manuscript and validated that it rendered without issues on our end. We'd like to apologize on any inconvenience caused during the reviewing process.

Sincerely,
The authors

Thank you for the reply, Fig. 1 was an online rendering issue only on openreview. It can be checked clearly. There are no more comments for me, I will keep the positive score as 3

Reviewer_oZL7OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Fast-TD-MPC, a framework for reducing the inference cost of data-driven model predictive control by adaptively deciding when planning is needed. The method is inspired by the dual-process theory of human cognition, pairing TD-MPC2's original MPPI planner, which is slow but robust to out-of-distribution observations, with a compact, efficient policy trained to imitate the planner. A lightweight detector routes between the two systems, so the fast policy handles familiar states and the planner is invoked only when needed. Since the method leaves the underlying world model and planner unchanged, it acts as a simple add-on that could be applied to existing TD-MPC2 agents. The paper evaluates this approach on 103 tasks across four domains, reporting significant gains in inference efficiency.

Critical strengths
  1. The per-step cost of sampling-based planning is a real obstacle to deploying methods like TD-MPC2, and the paper asks a sensible question about whether every decision needs the full planner. The dual-process framing makes the idea intuitive, and taking inspiration from the theory of human cognition is interesting. The method design is well-motivated, simple, and modular. Fast-TD-MPC leaves the world model and planner untouched and adds only a small imitation policy and a cheap routing gate. This makes the method easy to apply to existing agents, and the routing decision is a single threshold comparison, which can be applied to other similar methods.
  2. On DMControl, gating outperforms Round-robin at every matched latency and retains 96% of TD-MPC2's return at about 4× lower latency. The planning rate also rises substantially under disturbances (from 18% to 45% on DMControl), showing the gate responds to perturbed states. The empirical evaluation is extensive, with results reported on 103 tasks across four diverse domains with a single training configuration with three seeds per task.
Critical weaknesses
  1. Section 3 describes System 1 as being trained on states visited by the planner, but the reference distribution for the gate is built by rolling out System 1 itself. In the "theoretical" mode, every state System 1 visits is labeled in-distribution, but this includes states it reached by drifting away from the planner's distribution, where it may become less reliable. If this is a deliberate design choice, it would be helpful to give a clear explanation, as fitting the gate on the planner's own state distribution (System 1's training data) would seem more consistent.

    a. Some results suggest that the gate does not reliably identify when planning is needed in every setting. One example is on MyoSuite, where Round-robin at 10% recovers TD-MPC2’s success rate at 6.6 ms, but the gate still routes to the planner on 60% of steps at τp90\tau_{p90} (Table 12). Another example is ManiSkill2 at τp90\tau_{p90}, where the planner is called on only 13% of steps, even though System 1 alone achieves only 19% success against TD-MPC2’s 65% (Table 13).

    b. Since τp50\tau_{p50} is defined as the median of a set of held-out validation rollouts from System 1, it seems that approximately half of System 1’s own states will be flagged by construction. This might help explain the high planner rates at τp50\tau_{p50}. For example, at τp50\tau_{p50}, where Fast-TD-MPC tracks TD-MPC2 most closely under disturbance (Table 3), the gate routes to the planner on most steps (Tables 18–21), so this result largely reflects running the planner. On Meta-World, this configuration is slightly slower than baseline TD-MPC2 (27.2 vs. 26.8 ms shown in Tables 3 and 19).

  2. The efficiency gains are less general than presented in the paper’s introduction, and the closest existing methods are not compared.

    a. The introduction motivates Fast-TD-MPC by the control frequencies required for real-time robotics and by stale-action errors. Whenever the gate invokes MPPI, that step incurs the full planning latency (visible in Figure 5). The method therefore reduces average latency but not worst-case latency, which is what determines the achievable control frequency.

    b. The ~4x increase in inference efficiency is for DMControl only, and the ~2.7x average speedup figure includes a domain where performance drops by half (ManiSkill2, drop from 65% to 32% shown in Table 1).

    c. Appendix L describes AdaReP (Cheng et al., 2026) and speculative execution for TD-MPC2 (Lin et al., 2025b) as some of the most comparable approaches, since both skip planning conditionally. Neither of these is included as a baseline approach for comparison.

Questions for authors
  1. Reducing the planner from 512 to 64 samples barely changes latency (27.3 vs. 26.9 ms on DMControl, Table 1). Could the latency be dominated by per-iteration overhead? Was the planner compiled or otherwise optimized?
  2. TD-MPC2's policy prior already provides a fast reactive policy, and on ManiSkill2 it substantially outperforms System 1 (48% vs. 19%, Table 1), beating several Fast-TD-MPC configurations at lower latency. Did you evaluate gating between MPPI and the policy prior, and what motivates training a separate imitation policy?
  3. System 2 relies on a world model and value function that are trained only on the agent’s replay data. How reliable is the planner on states far outside that data?
5 comments
AuthorsOpenReview ↗

Response to Question Item 1-7

Thanks for the insightful comments.

Q1: Gate reference can include unreliable states from S1

We'd like to clarify that planner rollouts provide action targets for S1, while S1 define the fast policy's reference behavior. We note that expert and visitation can differ [1]. We then use this reference to detect deviations from the states that S1 encounters. Since it can include unsuccessful behavior, our reward-gated mode retains states whose per-step reward or completion proxy meets or exceeds the expert’s time-matched average.

Q2: Robustness at the median threshold relies on frequent planning

We’d like to clarify that τp50\tau_{p50} represents the conservative end of the performance–latency trade-off. It flags about half the held-out S1 calibration states by construction. Under disturbances, planning occurs on most steps, so its close performance to TD-MPC2 comes with a similar computational cost. On disturbed Meta-World, 99% planning gives 27.2 ms versus 26.8 ms for TD-MPC2. This point demonstrates fallback toward full planning, rather than an efficiency gain.

The evidence for preserving performance at lower cost comes from τp90\tau_{p90}. On disturbed DMControl, it retains 95% of planner return at half its mean step latency while planning on 45% of steps. We'll distinguish these two results in Sec. 4.3.

Q3: Gate planning choices are unreliable on MyoSuite and ManiSkill2; headline speedups overstates performance preservation across domains.

We note that the trade-off differs by domain. On MyoSuite, RR10% reaches 40% success at 6.6 ms, while all-p90 uses 60% planning for 37%. Reward filtering improves the gate’s trade-off to 39% at 15.7 ms, versus 39% at 31.4 ms for all-p50.

On ManiSkill2, p90 uses 13% planning and achieves 32% success versus the planner’s 65%; the policy prior reaches 48% at 11.1 ms. We will state these differences in Section 4.2 and pair each speedup with its performance. The 4.2x result concerns DMControl; the 2.7x average includes the ManiSkill2 performance loss.

Q4: Lower average latency does not establish a higher control frequency

We’d like to clarify that the reported speedup concerns average step latency. At τp90\tau_{p90}, 82% of clean DMControl steps avoid planning, reducing mean step time from 27.3 to 6.5 ms. Fallback steps retain the full planner cost plus routing overhead.

Our evaluation is synchronous, with the environment advancing after the action is computed. It measures execution time without testing how the environment changing during computation affects control. The results therefore establish lower average cost, rather than guaranteed deadline compliance or fewer stale-action errors.

Q5: The closest adaptive planning baselines are missing.

We’d like to clarify how our fast path differs from these approaches. AdaReP reuses cached plans [2], while Lin et al. correct queued actions using observation feedback [3]. Our S1 generates an action from the current latent without relying on a cached action sequence. The gate compares that latent against an offline reference distribution rather than a predicted state. We'll expand Appx. L with a qualitative comparison of the fast action generator, routing signal and training requirements. This will make the contribution and its relationship to existing adaptive planning methods clearer.

Q6: Why does reducing planner samples barely reduce latency, and was MPPI optimized?

We clarify that planning-based comparisons use the same uncompiled MPPI implementation. Reducing samples preserves six refinement iterations, whereas skipping a call removes the whole optimization loop. The flat timing is consistent with substantial fixed or sequential overhead; the measurements do not isolate its contribution.

Planner optimization can be combined with routing, although faster planning can reduce the relative speedup.

Q7: Was gating the existing policy prior evaluated, and why train a separate imitation policy?

We evaluated the prior as a standalone baseline, while the gated experiments used S1. The prior uses critic-based learning with entropy regularization, whereas S1 imitates the planner’s executed actions.

S1 has higher clean mean performance on DMControl (623 versus 560), Meta-World (87% versus 82%), and MyoSuite (35% versus 18%). ManiSkill2 favors the prior (48% versus 19%). These results motivate planner imitation in three domains and identify the prior as a stronger fast-policy candidate on ManiSkill2.

AuthorsOpenReview ↗

Response to Question Item 8

Q8: How reliable is S2 outside the world model’s and value function’s training data?

We’d like to clarify that S2 retains TD-MPC2’s pretrained world model, critic and full MPPI search budget. When invoked, it optimizes actions from the current state rather than relying on S1’s imitation of the planner.

Our disturbance results support this fallback choice. On disturbed DMControl, full planning achieves 616 return versus S1’s 405, while Fast-TD-MPC recovers 586. This demonstrates the value of retaining the original planner under the tested shifts. Far outside the replay distribution, its reliability remains dependent on the accuracy of the learned dynamics and value estimates.

References:

[1] Ross et al. A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning. AISTATS'11.

[2] Cheng et al. AdaReP: Adaptive Re-Planning under Model Mismatch for Neural World-Model Predictive Control. ICANN'26.

[3] Lin et al. Accelerating Multi-modal LLM Gaming Performance via Input Prediction and Mishit Correction. arXiv'25.

AuthorsOpenReview ↗

Gentle Reminder

Dear Reviewer oZL7,

We would appreciate hearing your thoughts on our response and whether it has addressed your concerns. Please feel free to let us know if any points would benefit from further clarification.

Sincerely,
The authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal and Updates

We thank all ACs and reviewers for their effort in the reviewing process. We summarize the discussion and our key updates below.

Key Contributions

Our work introduces Fast-TD-MPC, a lightweight and adaptive test-time planning framework for data-driven MPC, that routes between a compact amortized policy (System 1) and a full MPPI planner (System 2) via a lightweight OOD detector. Applied to TD-MPC2, it preserves the pretrained world model and planner while achieving up to ∼4×\sim4\times inference speedup on DMControl and ∼2×\sim2\times on ManiSkill2 with <10% performance drop on 3 of 4 domains, using a single set of hyperparameters across 103 tasks.

Strengths Recognized by Reviewers

The reviewers recognize:

  1. the practical simplicity and modularity of the framework (Fbsj, wj9N, oZL7),
  2. the extensiveness of the empirical evaluation across 103 tasks, three seeds, threshold sweeps, and disturbance tests (Fbsj, wj9N),
  3. the transparent and auditable gating mechanism, reducing routing to a single scalar comparison (Fbsj), and
  4. the promising systems idea of fast–slow architectures for efficient test-time planning (wj9N).

Addressing Reviewers' Concerns

We have addressed all main concerns raised by the reviewers and updated the manuscript accordingly (marked in blue):

  1. Scoping claims across domains: The abstract, introduction, contributions, and conclusion now explicitly scope speedup claims to DMControl, state the ManiSkill2 performance drop (65% →\to 32%), and acknowledge that gating clearly beats fixed scheduling only on DMControl (answering oZL7; Sections 1, 4.2, 7).

  2. Distinguishing τp50\tau_{p50} and τp90\tau_{p90}: Section 4.3 now separates the conservative τp50\tau_{p50} result (fallback toward full planning) from the τp90\tau_{p90} efficiency result (95% return retention at half latency with 45% planning on DMControl) (answering oZL7; Section 4.3).

  3. Pairing speedups with performance: Section 4.2 now reports each speedup alongside its per-domain performance, including domain-specific losses (answering oZL7; Section 4.2).

  4. Reframing the control-frequency motivation: The introduction now clarifies that the demonstrated benefit is lower average execution cost, not guaranteed real-time deadline compliance, since fallback steps retain full MPPI latency (answering oZL7, Fbsj; Section 1).

  5. Expanding the adaptive planning comparison: Appendix L now provides a side-by-side comparison of Fast-TD-MPC, AdaReP, Lin et al., and When to Plan across the fast action generator, routing signal, and training requirements (answering oZL7, Fbsj, wj9N; Appendix L).

  6. MPPI compilation: A new appendix discusses the use of uncompiled MPPI across all comparisons and the effect of compilation on relative speedup (answering oZL7; Appendix H).

Reviewers Fbsj and wj9N have acknowledged our clarifications and maintained their positive scores. We thank the reviewers and the AC for their time.

Sincerely,
The authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.