- 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
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.
Thank you
Dear Reviewer wj9N,
Thank you for your insightful comments.
Sincerely,
The authors
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.