acceptodds
Under review as a conference paper at ICLR 2027

Map2Motion: Brain-Inspired Prospective Planning from Structural Cognitive Maps to Executable Motion

Abstract

Mammalian navigation suggests a computational regime in which compact structural knowledge is continuously grounded by prospective computation to support flexible long-horizon behavior. In contrast, conventional robotic navigation typically relies on dense metric representations, accurate localization, and globally specified trajectories, whose storage, computation, and metric-consistency requirements grow with environment scale. Inspired by structural cognitive maps and prospective neural sequences, we propose Map2Motion, a cognitive-map-guided prospective motion planning converting structural cognitive maps to executable motion. Map2Motion conditions a flow-based distribution of prospective motions on structural guidance and current perception, and learns a forward dynamics model to predict the generated motions' embodied future states and traversability for closed-loop selection and execution. Our experimental result on MP3D benchmark achieves an 83% success rate (SR) with only 4.7% of the global map storage of dense-map baselines, while retaining 64% SR under 0.2 m localization error. In real-world building navigation tasks, it achieves 78% SR with 19.32 ms average planning latency and only 2.6% of dense-map storage, indicating its efficiency and capability in practical scenarios.

Primary area:
applications to robotics, autonomy, planning
Keywords:
sparse topological navigation, brain-inspired prospective motion planning, structural cognitive map

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 2 reviews, above 85% of the others.
Reviewer_kJd6OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Map2Motion, a robot navigation framework that converts guidance from a sparse room–doorway graph into executable local controls. Motivated by structural cognitive maps and prospective planning, the method uses A* to extract a structural route and conditions a flow-matching model on this route, current depth observations, and recent motion context to generate multiple candidate control sequences. A learned forward dynamics model predicts future robot states and traversability for each candidate. These predictions support candidate selection, followed by execution and closed-loop replanning. Both learned components are trained using offline simulation rollouts. Experiments report navigation high success rates, alongside reduced global planning-map storage and measured planning latency relative to selected dense-map baselines. Additional experiments evaluate robustness to translational localization noise and the contributions of structural guidance, motion generation, and predictive evaluation.

Critical strengths
  1. The paper addresses the practical challenge of converting sparse waypoints into executable local motion. Its combination of structure-conditioned motion generation and conditioned forward prediction gives the components complementary roles; the generator proposes relevant control sequences, while the predictor estimates their outcomes before selection and execution.

  2. Experimental results report an 83.47% simulation success rate while using only 4.7% of the global planning-map storage of the dense-map baselines. The reported planning latency is also substantially lower than that of Dijkstra + MPPI. Although the strongest dense-map baseline achieves higher navigation success, these results demonstrate the potential of sparse structural representations for reducing planning costs while retaining useful performance.

Critical weaknesses
  1. The paper emphasizes the substantial reduction in global planning-map storage as a major advantage. However, the dense-map baselines require already relatively modest storage requirements for modern robotic platforms. The authors do not demonstrate that map storage constitutes a practical bottleneck, nor do they analyze how storage requirements scale with environment size. Furthermore, the reported storage comparison excludes the memory footprint of the learned motion generator and forward dynamics model. Consequently, the practical significance of the reported storage savings remains unclear.

  2. The FDM is trained using recorded simulation rollouts but is subsequently used to evaluate candidate control sequences that have not been executed. The paper does not investigate how well the model generalizes to such candidate actions, particularly potentially unsafe or out-of-distribution motions. Since traversability is trained using continuous TSDF-derived targets, the authors should report regression errors and, under clearly specified safety thresholds, precision and recall for identifying unsafe candidate trajectories.

Questions for authors

Could you report direct evaluations of the FDM, including future-position and orientation errors, clearance prediction errors, and precision/recall for detecting unsafe trajectories under clearly defined safety thresholds?

Why does increasing the number of candidate motions degrade navigation performance? At K=4, the success rate drops from 83.47% with 4 candidates to 71.90% with 128 candidates. Does this degradation arise from lower-quality proposals, inaccuracies in the FDM, or errors in the candidate-scoring mechanism?

2 comments

Thank you for the insightful and constructive comments, and for recognizing the complementary design of our method and its advantages in planning efficiency and map storage. We have carefully considered the concerns through additional experiments evaluating FDM prediction accuracy, investigating the impact of candidate count on navigation performance, and analyzing map-storage scalability. These results provide further insights into the effectiveness and limitations of our approach, and have been incorporated into the revised manuscript.

W2 and Q1. Direct evaluation of FDM predictions

We evaluated whether the FDM generalizes from recorded rollouts to CFM-generated motions in unseen scenes. We collected 2,266 planning snapshots from 159 randomly sampled episodes across five unseen scenes. At each snapshot, 128 candidates were generated, and FDM predictions were compared with simulated rollouts under the same controls.

Position error (m) ↓Orientation error (°) ↓Trajectory clearance MAE (m) ↓
0.0564.030.029
[0.049, 0.064][3.80, 4.25][0.025, 0.034]

We further evaluated unsafe-motion detection using rollout clearance below 0.45 m as the ground-truth unsafe threshold and predicted risk above 0.4 as the detection threshold.

Recall (%) ↑Precision (%)F1 (%)AP
65.1246.7034.100.9319

The FDM achieves a mean position error of 0.056 m, heading error of 4.03°, and clearance MAE of 0.029 m. It also achieves 65.12% recall in detecting unsafe motions. These results demonstrate that the FDM retains trajectory prediction and unsafe-motion detection capabilities on CFM-generated motions in unseen scenes.

Q2. navigation performance degradation arising from inaccuracies in the FDM

We appreciate the reviewer's insightful observation and thought-provoking hypothesis. We first examined whether FDM accuracy changes with candidate count, using the same 2,266 snapshots and candidate sets with N∈{4,8,16,32,64,128}N\in\{4,8,16,32,64,128\}.

NRecall (%)Precision (%)F1 (%)AP
465.5647.5134.540.9862
866.3047.2134.570.9640
1666.1546.9334.340.9544
3265.7946.6734.240.9458
6465.6146.6534.180.9362
12865.1246.7034.100.9319

Recall, precision, and F1 remain largely stable as NN increases suggesting that the FDM becomes less effective at distinguishing among candidates in larger pools. This may be attributed to the FDM's limited model capacity and the increased difficulty of discriminating among a larger number of motion candidates.

Navigation performance depends not only on the accuracy of individual predictions but also on whether these predictions lead to the selection of high-quality motions. Therefore, we further examine the quality of the selected motions using the following selection-level metrics.

NGT top-5 hit rate (%) ↑\uparrowTrue selected cost ↓\downarrowOracle best cost ↓\downarrowSelection regret ↓\downarrow
4100.0000.9500.9110.039
895.6000.8960.8440.051
1694.0000.8500.7880.062
3291.9000.8120.7400.072
6489.5000.7810.6990.082
12886.3000.7540.6620.092

As N increases, both oracle best cost (0.911→0.662) and FDM-selected cost (0.950→0.754) decrease, while selection regret increases (0.039→0.092) and GT top-5 hit rate decreases (100.0%→86.3%). These results indicate that the FDM becomes less effective at selecting the best candidates from larger pools, which may contribute to the observed decline in navigation success rates. Improving the FDM's candidate-ranking capability may help mitigate this degradation.

W1. Practical significance and scaling of map storage

We agree that dense occupancy maps may not pose a significant storage bottleneck in small and medium-sized environments on modern robots. However, dense metric representations incur increasing storage costs as environments scale, motivating the use of topological and hierarchical representations for large-scale spatial reasoning and long-term navigation (Hughes et al., IJRR 2024; Vutetakis and Xiao, IJRR 2025; Cheng et al., IEEE T-RO 2025).

This also has a biological motivation: cognitive maps support flexible navigation through abstract spatial relationships rather than requiring detailed metric representations at every stage (Tolman, 1948). Dense metric maps represent spatial details, whereas sparse topological maps capture structural relationships. Map2Motion uses this structural abstraction to guide predictive motion generation and produce executable navigation plans.

To quantify the storage advantage, we compared dense occupancy maps and sparse topological maps across eighteen Matterport3D scenes of different sizes.

SceneEnvironment area (m²)Dense occupancy-map (KiB)Sparse topological-map (KiB)
HxpKQynjfin56.0522.0117.83
17DRP5sb8fy136.1253.2925.46
zsNo4HB9uLZ299.71117.2051.73
q9vSo1VnCiC502.79196.5260.90
ZMojNkEp431698.06272.8048.11
Z6MFQCViBuw1228.35479.9552.16
dhjEzFoUFzH2408.25940.849.71
B6ByNegPMKs4912.321918.99192.51

The results show that sparse topological maps require less storage across all evaluated scenes, with increasingly larger savings in larger environments. Unlike dense occupancy maps, topological-map storage primarily depends on the number of nodes and their connectivity.

Manuscript revisions

Thank you for the thoughtful comments of our work, we have made these revisions below following your suggestions.

  1. Added map-storage scaling curves to the appendix Fig.6C.
Reviewer_P4PvOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

As I understand it, given a sparse structural graph, this paper first performs global planning with A* to generate a structural route PtP_t. Then, Conditional Flow Matching (CFM) generates multiple candidate control sequences conditioned on the structural route, local depths, and recent state and control history. A learned FDM model predicts the future states and traversability for each candidate, which are evaluated using predefined costs (progress, safety, smoothness). The main contribution of this paper is a structure-to-motion planning framework that transforms sparse topological guidance into executable local motion through generative motion proposals and learned predictive evaluation, without requiring a dense global geometric map.

Critical strengths

S1. Proper problem formulation and intuitive framework. I think the main question this paper addresses is relevant and meaningful: How can a robot navigate using structural relations without relying on dense geometric reconstruction? The proposed framework uses sparse structural information for long-range guidance while relying on local observations for executable motion planning. I find this separation intuitive and well-motivated, especially for large-scale navigation where maintaining dense geometric maps can be expensive. The combination of generative motion planning and predictive evaluation also provides a reasonable way to bridge abstract structural guidance and low-level robot control.

S2. Strong experimental results and practicality. The results are convincing, especially considering the limited global map info used. It achieves reasonable success rate compared to the dense-map baseline while requiring only less than 5% of it map storage and lower planning latency. The real-world experiemnts in a large indoor environment is also a critical point to be appreciated.

Critical weaknesses

W1. Limited analysis of generalization for the generator and FDM. My main concern is how well the learned policy and FDM generalize to unseen environments and motion distributions. Both are trained offline with demo rollouts, but the paper does not sufficiently clarify the generalization protocol or analyze performance under diverse unseen configs. In particular, the FDM evaluators control sequences, which may differ from the training distributions. This could directly affect action selection and navigation safety.

W2. Insufficient ablation and analysis of individual design choices. Although the paper provides ablations for the major components, I would like to better understand the contribution of some specific design choices. For example, both the current robot state and historical motion information are used as inputs, but it is unclear how much additional benefit the historical context provides.

Questions for authors

Q1. Could the authors clarify the training/test scene split and show how well the method generalizes to unseen environments?

Q2. How accurately does the FDM predict future trajectories and traversability for CFM-generated control sequences, rather than demonstration data? Could prediction errors explain the performance drop as N increases?

Q3. What is the actual contribution of state and control history beyond the current observation and robot state? A simple ablation would be helpful.

Q4. Could the authors compare against a topology-guided MPPI baseline with the same sparse graph and depth observations? For example, the next doorway could serve as a subgoal with a local depth-based costmap.

4 comments

Thank you for the constructive feedback and positive assessment of our problem formulation, the well-motivated framework of topological motion planning, and the promising experimental results in terms of navigation performance, map efficiency, planning latency, and real-world applicability. Following your constructive suggestions, we have clarified the generalization protocol and conducted additional experiments on FDM prediction accuracy, topology-guided MPPI comparison, and motion history ablations. The results further demonstrate the effectiveness of our framework, achieving 86.94% success on unseen scenes and outperforming topology-guided MPPI, while providing deeper investigations into the contributions of individual components. We detail our responses and corresponding revisions below.

Q1. Generalization to unseen environments

We trained Map2Motion on nine Matterport3D scenes and evaluated it on five held-out scenes. For each test scene, we randomly sampled 121 episodes, resulting in 605 evaluation episodes.

Test sceneEpisodesSR (successes/episodes)95% CI (SR)SPL
B6ByNegPMKs12186.78% (105/121)[79.60%, 91.69%]0.7899
Pm6F8kyY3z212193.39% (113/121)[87.49%, 96.61%]0.8730
ZMojNkEp43112160.33% (73/121)[51.43%, 68.60%]0.5168
r47D5H71a5s12195.04% (115/121)[89.60%, 97.71%]0.8240
x8F5xyUWy9e12199.17% (120/121)[95.47%, 99.85%]0.9151
Overall60586.94% (526/605)[84.02%, 89.40%]0.7837

Map2Motion achieves an overall success rate of 86.94% and SPL of 0.7837 across the five unseen scenes. Although performance varies with scene complexity, it maintains a success rate of 60.33% even in the challenging open scene ZMojNkEp431, which contains numerous cluttered obstacles.

We further evaluated the system in real-world environments with different regions, lighting conditions, sensor noise, and dynamic obstacles. Map2Motion achieves a 78.3% success rate, further demonstrating its transfer to real-world deployment conditions. We have added the environment split and evaluation protocol to the revised manuscript.

W1 and Q2. FDM prediction on generated controls and its effect on selection

Thank you for raising this insightful question and suggesting a plausible explanation for the observed phenomenon. We tested the FDM's generalization performance both on prediction accuracy and its effect on candidate selection, as detailed below.

W1 & Q2. Trajectory prediction error on generated controls

We collected 2,266 planning snapshots at 1-second intervals from 159 randomly sampled episodes across five unseen scenes. At each snapshot, we generated 128 CFM candidates and compared the FDM-predicted trajectories against ground-truth simulated rollouts under the same control sequences.

For each metric, we first averaged the prediction errors across the 128 candidates at each snapshot, then across snapshots within each episode, and finally across episodes with equal weight. This episode-balanced aggregation prevents episodes with more planning snapshots from disproportionately influencing the reported results. We report 95% confidence intervals estimated using episode-level bootstrap resampling, preserving all snapshots and candidates within each sampled episode.

Position error(m)Orientation error(°)Trajectory clearance MAE(m)
0.056 [0.049, 0.064]4.03 [3.80, 4.25]0.029 [0.025, 0.034]

On CFM-generated controls in unseen environments, the FDM achieves a mean position error of 0.056 m, a heading error of 4.03°, and a trajectory clearance MAE of 0.029 m. These results demonstrate that the FDM retains useful trajectory and clearance prediction capability for CFM-generated motions in unseen scenes.

Q2. Could prediction errors explain the performance drop as N increases?

We evaluated FDM prediction and candidate-selection performance across N∈{4,8,16,32,64,128}N\in\{4,8,16,32,64,128\} using 2,266 planning snapshots and simulated rollouts as ground truth. Unsafe motions are defined as those with rollout clearance below 0.45 m, and predicted risk above 0.4 indicates an unsafe prediction. We use recall to evaluate unsafe-motion detection and additional metrics to assess candidate quality and selection performance.

NRecall (%) ↑GT top-5 hit rate (%) ↑FDM-selected cost ↓Oracle best cost ↓Selection regret ↓
465.56100.00*0.9500.9110.039
866.3095.600.8960.8440.051
1666.1594.000.8500.7880.062
3265.7991.900.8120.7400.072
6465.6189.500.7810.6990.082
12865.1286.300.7540.6620.092

The results show three trends:

  1. Unsafe-motion detection: Recall peaks at N=8 (66.30%) and remains relatively stable across N.
  2. Candidate quality: Both oracle best cost (0.911→0.662) and FDM-selected cost (0.950→0.754) decrease as N increases, indicating improved candidate availability.
  3. Candidate selection: Selection regret increases (0.039→0.092), while GT top-5 hit rate decreases (95.6%→86.3% for N≥8), indicating reduced selection effectiveness with larger candidate pools.

Overall, N=8 provides the best trade-off, followed by N=16. Larger candidate pools offer better potential motions, but the FDM cannot fully exploit them due to its limited candidate-selection capability. This may contribute to the observed decline in navigation success.

W2 and Q4. Topology-guided MPPI baseline

We implemented topology-guided MPPI using the next doorway on the topological route as a subgoal and a local depth-based costmap to evaluate sampled motions. Both methods were evaluated on the same 605 episodes across five held-out scenes.

MethodOverall SRSPLSR, reference path >5 mSR, route with >1 doorwayMean completion time (s)
Topology-guided MPPI62.81% (380/605)0.563330.80% (69/224)12.16% (9/74)20.17
Map2Motion (ours)86.94% (526/605)0.783768.30% (153/224)50.00% (37/74)9.72

The topology-guided MPPI baseline encountered no feasible motion candidates in at least one planning step in 84.89% of failed episodes. Additionally, 102 failed episodes involved no collisions, while 55 timeout episodes involved collisions or persistent obstruction despite feasible plans being available.

The key difference is how topological guidance is converted into local motion. MPPI uses the next doorway as a subgoal and evaluates sampled kinematic rollouts using a local costmap. In contrast, Map2Motion conditions motion generation on structural-route context and predicts candidate outcomes using perception, motion history, and control inputs.

Overall, Map2Motion improves success rate from 62.81% to 86.94% and reduces mean completion time from 20.17 s to 9.72 s on jointly successful episodes, demonstrating the advantage of the complete motion-generation and prediction framework over topology-guided MPPI.

W3 and Q3. Contribution of state and control history

Thank you for the thought-provoking question about the contribution of history input. We removed the history inputs from CFM or FDM separately and retrained the corresponding model. We evaluated them on five unseen scenes with 605 randomly-sampled tasks.

VariantOverall SRSR, tasks ≥5 mSR, tasks with ≥1 doorwaySR, tasks with ≥2 doorwaysCollision-flagged steps, ≥5 m tasks (%)Mean recovery time, ≥5 m tasks (s)
CFM without history84.13% (509/605)60.71% (136/224)61.76% (105/170)36.49% (27/74)29.052.017
FDM without history70.58% (427/605)44.20% (99/224)44.12% (75/170)24.32% (18/74)45.942.774
Full model86.94% (526/605)68.30% (153/224)68.24% (116/170)50.00% (37/74)27.461.296

Incorporating historical states improves performance particularly on long-range navigation tasks, as temporal context helps capture motion trends and supports consistent decisions over extended horizons, as reflected in the higher success rates on tasks exceeding 5 m (68.30% versus 60.71% and 44.20%). Historical information also contributes to navigation stability and recovery efficiency by providing context about recent motion dynamics, resulting in fewer collision-flagged steps (27.46% versus 29.05% and 45.94%) and shorter recovery times (1.296 s versus 2.017 s and 2.774 s). These results suggest that historical states provide valuable temporal context for both trajectory prediction in FDM and motion generation in CFM, improving navigation robustness.

I appreciate the authors for their detailed response. Most of the responses are on point and sufficient to address my questions. Additional experiments showed strong performances even in unseen environments, and the comparison with topology-MPPI is also valid. I don't find any further concerns with this paper for now. Therefore, I keep my positive rating.

We sincerely appreciate your thoughtful follow-up and positive assessment of our work. We are glad that our responses have adequately addressed your concerns, and we will further refine and polish the manuscript according to your constructive suggestions to make it as strong as possible.

Comments on the paper

6 comments

Thank you for your careful review and for recognizing the value of our structural-to-motion interface for navigation, its promising storage–performance trade-off, and the complementary evidence from ablations and real-world deployment. We have carefully considered all your concerns and and clarified the evaluation protocol, confirming 86.94% success across 605 episodes in five held-out scenes, conducting additional experiments on structural-route conditioning, FDM candidate ranking, and investigating our Map2Motion's robustness to correlated localization drift. We have also corrected the identified unclear point and incorporated these findings into the revised manuscript. We address each comment in detail below.

Q1 and W1. Scene-level generalization and evaluation protocol

Thank you for pointing out the missing statements about scene split and evaluation denominators, and giving us an opportunity to clarify our generalization capabilities. Our training and test scenes are disjoint, as detailed below.

We trained on nine multi-room Matterport3D scenes and evaluated on five held-out scenes.

SplitScene IDs
Training17DRP5sb8fy, 2t7WUuJeko7, JeFG25nYj2p, mJXqzFtmKg4, q9vSo1VnCiC, RPmz2sHmrrY, ur6pFq6Qu1A, Z6MFQCViBuw, zsNo4HB9uLZ
TestB6ByNegPMKs, r47D5H71a5s, ZMojNkEp431, Pm6F8kyY3z2, x8F5xyUWy9e

For the navigation evaluation in §5.1 and the module ablations in §5.5, we randomly sampled 121 episodes per test scene, yielding 605 episodes in total. Following further refinement during the discussion period, the updated Map2Motion results are given below.

Test sceneEpisodesSR (successes/episodes)95% CI (SR)SPL
B6ByNegPMKs12186.78% (105/121)[79.60%, 91.69%]0.7899
Pm6F8kyY3z212193.39% (113/121)[87.49%, 96.61%]0.8730
ZMojNkEp43112160.33% (73/121)[51.43%, 68.60%]0.5168
r47D5H71a5s12195.04% (115/121)[89.60%, 97.71%]0.8240
x8F5xyUWy9e12199.17% (120/121)[95.47%, 99.85%]0.9151
Overall60586.94% (526/605)[84.02%, 89.40%]0.7837

Map2Motion achieves an overall success rate of 86.94% and SPL of 0.7837 across the five unseen scenes. Although performance varies with scene complexity, it maintains a success rate of 60.33% even in the challenging open scene ZMojNkEp431, which contains numerous cluttered obstacles.

The computational evaluation in §5.2 and localization-noise experiments in §5.3 used 121 randomly sampled episodes in the held-out scene B6ByNegPMKs. The real-world evaluation in §5.4 comprised 46 trials with randomly selected goals.

We have added these protocol details and the real-world trajectory distribution to the revised manuscript. The held-out scene evaluation, with an overall 86% success rate, supports transfer to unseen simulated environments. The physical experiments also involve regions, lighting conditions, sensor noise, and dynamic obstacles that differ from the training data, providing further evidence of applicability beyond the training sets.

W2. Isolating structural-route conditioning

Following the reviewer's suggestion, we conducted a matched comparison by retraining a single-point variant conditioned only on the next topological node, while retaining the same planning framework. Both variants were evaluated on the same 605 navigation tasks across five unseen scenes.

VariantSPL, ≥5 m tasks (%) ↑SPL, ≥1 doorway (%) ↑Collision-flagged steps (%) ↓Completion time (s) ↓
Single-point conditioning54.3050.3230.5113.25
Structural-route conditioning60.4658.5927.4612.11

Structural-route conditioning improves SPL by 6.16 percentage points on long-distance tasks and 8.27 points on doorway-traversal tasks, while reducing collision-flagged steps and completion time. Failure analysis also shows fewer persistent obstructions after recovery attempts.

We further analyzed 27 topological transition events across 17 tasks, using shortest paths from dense occupancy costmaps as references.

VariantHeading error at switching (°) ↓Mean heading error during next 1 s (°) ↓Alignment time (s) ↓
Single-point conditioning30.9627.200.733
Structural-route conditioning25.9523.980.668

Structural-route conditioning reduces heading error at transitions and during the following second, and achieves stable alignment faster. These results indicate that downstream route context improves local motion selection and alignment with upcoming path segments.

W2 and Q2. FDM continuation and candidate selection

Continuation beyond the conditioned action horizon

Beyond the supplied control horizon, the FDM predicts subsequent motion using learned dynamics and contextual information. Inspired by multiscale predictive horizons and goal-directed hippocampal sequences (Brunec et al., 2022; Pfeiffer et al., 2013), short control sequences constrain near-term motion, while the FDM predicts longer-horizon trajectories using spatial structure, observations, and learned behavioral priors.

We train the FDM using demonstrated trajectories extending beyond the control horizon, allowing it to learn the relationship between short-term controls and longer-term motion. The validation position MSE converges to 0.09, indicating that the model learns to predict trajectories beyond the supplied control horizon.

Ranking and selection under generated actions

We evaluated FDM ranking accuracy by comparing its scores for CFM-generated candidates against ground-truth (GT) rollout scores, using the same scoring function. We collected 2,266 planning snapshots from 159 episodes across five unseen scenes, evaluating 128 candidates per snapshot.

Spearman's ρ ↑Pairwise agreement (%) ↑GT top-5 hit rate (%) ↑Selection regret ↓
0.81984.186.30.092

The FDM achieves a Spearman correlation of 0.819 and 84.1% pairwise agreement with GT rankings. Its selected candidate falls within the GT top five in 86.3% of snapshots, with a selection regret of 0.092. These results demonstrate that the FDM can effectively rank and select high-quality motions from 128 generated candidates.

Suggested revision. Correlated localization drift in §5.3

Following your constructive recommendation, we added an evaluation of temporally correlated localization error. We use a Gauss–Markov process to model persistent drift in the localization estimate supplied to the planner.

At the start of each episode, we sample localization drift error $ \mathbf b_0\sim\mathcal N(\mathbf 0,\sigma_{xy}^{2}I_2). $ At each planning update, with Δt=0.05\Delta t=0.05 s, we evolve the state estimateion bias as $ \mathbf b_{j+1} =\alpha\mathbf b_j +\sigma_{xy}\sqrt{1-\alpha^2},\mathbf z_j, \mathbf z_j\sim\mathcal N(\mathbf 0,I_2), \alpha=e^{-\Delta t/\tau}. $ The parameter σxy\sigma_{xy} sets the per-axis error standard deviation, while τ\tau controls its temporal persistence. We add the current bias to the planner's localization input and evaluate 80 episodes per condition. The results of success rate are as follow.

Correlation time τ\tau (s)Methodσxy=0.1\sigma_{xy}=0.1 mσxy=0.2\sigma_{xy}=0.2 mσxy=0.3\sigma_{xy}=0.3 mσxy=0.4\sigma_{xy}=0.4 m
2Map2Motion73.75%71.25%81.25%78.75%
2Dijkstra+MPPI71.25%73.75%81.25%80.00%
10Map2Motion77.50%82.50%78.75%72.50%
10Dijkstra+MPPI82.50%71.25%45.00%23.75%

The advantage of Map2Motion is more pronounced under persistent drift of larger amplitude: at τ=10\tau=10 s and σxy=0.3\sigma_{xy}=0.3 and 0.4 m, its SR is 78.75% and 72.50%, exceeding Dijkstra+MPPI by 33.75 and 48.75 percentage points. These results support robustness to temporally correlated localization errors in the tested conditions, particularly sustained positional biases of larger amplitude.

Reply to additional comments

Temporal configurations in §5 and the appendix

Thank you for catching this inconsistency. We have corrected the appendix to match the temporal configuration used in the experiments, as summarized below.

ComponentSequenceStepsintervalNominal horizon
CFMMotion history80.05 s0.4 s
CFMGenerated candidate controls160.05 s0.8 s
FDMState history160.5 s8 s
FDMPredicted future states80.5 s4 s

Latency statistics in §5.4

We thank the reviewer for pointing out that the latency statistics were mislabeled. We have corrected §5.4: 19.321 and 221.584 ms are the medians for Map2Motion and the real-world baseline, respectively; the corresponding means are 21.037 and 368.440 ms.

Executable implementation

Thank you for your interest in our work's implementation. We plan to release the code upon acceptance. To support reproducibility during review, we will provide the implementation and configuration details as fully as possible.

Reply to additional comments

Risk loss in Eq. (14)

Thank you for pointing out the missing risk-loss definition. We have expanded the appendix to specify the loss and its training targets. For each future trajectory, we compute the clearance and convert it to a sigmoid-based soft risk label. The risk loss is binary cross-entropy between the predicted risk logits and these targets.

Specifically, we divide the ground-truth trajectory into consecutive 0.5 s prediction intervals starting at the current pose and define chc_h as the minimum TSDF value over footprint samples at interpolated poses within the hh-th interval. We then define the segment-wise soft risk and cumulative target as

=\sigma\!\left(\frac{d_{\mathrm{safe}}-c_h}{\tau_r}\right), \qquad \rho_h^{\mathrm{gt}} =\max_{1\le j\le h}\rho_j^{\mathrm{seg}},$$ where $\sigma$ denotes the sigmoid function, $d_{\mathrm{safe}}=0.45$ m, and $\tau_r=0.07$ m. During target construction, the sigmoid argument is clipped to $[-60,60]$ for numerical stability. The cumulative target represents the highest sampled geometric risk encountered up to future step $h$. We use cumulative targets to preserve earlier high-risk encounters in the supervision, even when the candidate trajectory subsequently moves into a safer region. Given predicted risk logits $z_{h}$ and $\widehat\rho_{h}=\sigma(z_{h})$, the risk loss is $$L_\rho= -\frac{1}{L_s} \sum_{h=1}^{L_s} \left[ \rho_{h}^{\mathrm{gt}}\log\widehat\rho_{h} +(1-\rho_{h}^{\mathrm{gt}}) \log(1-\widehat\rho_{h}) \right],$$ where $B$ is the batch size and $L_s$ is the number of predicted future steps. We implement this loss directly on logits using numerically stable binary cross-entropy and include it in the total objective with weight $\lambda_\rho=2.0$. ## Manuscript revisions Thank you for the careful examination of our work, we have made these revisions below following your suggestions. 1. Clarified the experimental protocol and generalization evaluation in §5 and §5.1. 2. Added the matched structural-route conditioning comparison in §5.5. 3. Added correlated-localization-drift experiments in §5.3. 4. Corrected the latency terminology in §5.4. 6. Reconciled the temporal configurations in the appendix. 7. Reconciled the risk-target construction and risk loss in the appendix.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.