acceptodds
Under review as a conference paper at ICLR 2027

A Transport-Robot Navigation Platform for Pre-Collision Path Diagnosis and Repair

Abstract

Transport robots have been widely deployed in automated transportation scenarios such as warehouse logistics. When executing preplanned paths, transport robots may still collide with environmental obstacles, causing payload loss and subsequent task interruption. Existing methods mainly focus on post-collision recovery or replanning, without dedicated methods or data support for pre-collision path diagnosis and repair. To address this gap, we introduce RobotPath, a transport-robot navigation platform for pre-collision path diagnosis and repair in warehouse environments. RobotPath integrates three components: (1) scene simulation and automatic path generation, supporting path generation and execution across different robot configurations, shelf layouts, and execution regimes; (2) repair generation and verification tools, which automatically generate repairs and verify their geometric feasibility; and (3) LLM-RobotPath, a platform-native path diagnosis and repair model using numerically aware collision diagnosis and verifier-guided repair. We further release RobotPath-Bench, the first benchmark for this task, containing 20,399 path-execution records across 5 robot configurations, 7 shelf layouts, and 2 execution regimes. Extensive experiments show that RobotPath and RobotPath-Bench provide reliable support for modeling and evaluating pre-collision path diagnosis and repair. Our code and dataset are available at URL.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Transport Robot, Collision Diagnosis, Path Repair, Embodied Platform

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

The paper introduces RobotPath, a simulation platform for diagnosing and repairing collision-prone warehouse robot paths. RobotPath-Bench contains 20,399 execution records across five robot configurations, seven layouts, and two control regimes. LLM-RobotPath uses structured geometric inputs for collision diagnosis and generates local repairs through supervised learning and GRPO, with geometric verification and planner fallback. Experiments evaluate diagnosis, local repair, configuration transfer, and verifier agreement with simulated execution.

Critical strengths
  1. Sections 3.1–3.2 provide collision labels, object and segment identification, contact localization, and separate robot/layout transfer settings. Table 2 demonstrates that these metrics reveal failures hidden by collision classification alone.
  2. Section 4.2 independently executes 1,213 geometrically verified repairs in simulation, with 1,202 successful arrivals. This provides useful evidence for the reliability of the platform’s repair-generation and verification tools.
  3. Table 3 distinguishes greedy repair, best-of-eight sampling, fallback frequency, and final success. GRPO improves greedy success from 68.00% to 92.50% on the evaluated contexts. The platform–dataset–model organization is understandable.
Critical weaknesses
  1. Tables 1–3 compare only variants of LLM-RobotPath. No geometric collision-checking baseline, standalone planner, or alternative learned model is evaluated. Since inputs include explicit robot and obstacle geometry and the platform already implements Hybrid A* and deterministic fallback, matched comparisons are important for establishing benchmark difficulty and the learned model’s practical value.
  2. The novelty discussion needs a clearer relationship to established pre-execution collision checking and trajectory refinement, including FCL (Pan et al., 2012) and CHOMP (Ratliff et al., 2009). These do not invalidate the dedicated execution-annotated benchmark, but they limit the broader characterization of prior work as primarily intervening during or after execution.
  3. Repair evaluation is restricted to valid Simple-collection contexts and depends strongly on target adjustment. Figure 5 reports context coverage of 97.32% at a 3 m displacement allowance, 58.15% at 0.25 m, and 1.95% with no displacement. It is also unclear whether Table 3 uses predicted diagnosis throughout or whether the complete path continuation is verified after target adjustment. Section 4.2’s 99.09% result evaluates platform-generated repairs rather than the complete learned pipeline.
  4. Table 3 identifies its GRPO final-success result as separately selected without specifying the selection criterion. The supplied PDF also omits Appendices A–D, preventing assessment of details deferred there; their supplementary availability needs clarification.
Questions for authors
  1. Are standalone geometric-checking or planner results available under matched constraints?
  2. Does Table 3 use predicted collision information, and how was its separately reported GRPO result selected?
  3. After target adjustment, is continuation along the remaining original path verified and executed?
  4. Are Appendices A–D available separately?
3 comments

We thank the reviewer for the valuable comments and address each point as follows:

W1 & Q1: Standalone geometric-checking or planner results under matched constraints.

We use the swept-footprint collision-checking procedure underlying the Geometric Verifier (Sec. 2.2), together with controller-kinematic replay, as our non-LLM collision-diagnosis baseline. This allows us to directly evaluate geometric collision checking on the diagnosis test sets. We further introduce two categories of baselines: (1) two LLMs (GPT-6-Astra and DeepSeek-flash v3); and (2) lightweight MLP- and GNN-based models. Specifically, the MLP processes flattened geometric features, while the GNN employs multi-head self-attention over a fully connected graph of geometric entities, followed by graph-level pooling to obtain a scene representation. Both models jointly predict collision occurrence, the colliding object, the first collision segment, and the contact location. The comparison results are presented below. All metrics are percentages except MCE, which is reported in meters.

ModelBAcc ↑F1 ↑ObjAcc ↑SegAcc ↑[email protected] ↑[email protected] ↑MCE ↓
Simple collection
Geometric Verifier95.6397.8699.7681.5135.0471.780.1723
MLP66.1382.3960.3423.841.954.871.0190
GNN91.1495.2898.0577.8614.3642.580.2640
GPT-6-astra93.9295.0192.7070.8053.7785.400.1097
DeepSeek-flash v395.6396.6895.6272.2658.3988.560.1085
LLM-RobotPath98.4299.0398.7890.5179.8191.480.0820
Hard collection
Geometric Verifier50.5281.5970.5348.123.617.052.2647
MLP67.6385.2732.762.660.160.315.6120
GNN59.8442.057.050.310.000.005.1180
GPT-6-astra48.1578.7664.2650.003.613.611.7938
DeepSeek-flash v367.4185.5160.3439.661.725.172.0199
LLM-RobotPath93.6694.9388.0976.1864.2668.501.0073

For the non-LLM local-replanning baseline, we introduce the Fallback Planner (FP), a standalone deterministic planning method. SFT and GRPO refer to supervised fine-tuning and subsequent GRPO optimization, respectively. Greedy (Top-1) uses a single greedily generated candidate, while Bo8 selects from eight sampled candidates. w/ FP invokes the fallback planner when no model-generated candidate passes geometric verification. The evaluation uses 411 collision cases from the Simple test set, of which 400 have valid repair contexts. All methods in this repair comparison use ground-truth collision context, identical repair constraints, and the same geometric verification procedure.

MethodLocal / 400All collisions / 411Fallback addedMedian generation/search (s)
Fallback Planner (FP)388/400 (97.00%)94.40%00.1252
SFT_Greedy289/400 (72.25%)70.32%00.9058
SFT_Bo8380/400 (95.00%)92.46%02.5720
SFT_Bo8 w/ FP396/400 (99.00%)96.35%162.5720
GRPO_Greedy375/400 (93.75%)91.24%00.8281
GRPO_Bo8385/400 (96.25%)93.67%01.1328
GRPO_Bo8 w/ FP397/400 (99.25%)96.59%121.1328

The two tables report collision-diagnosis and local-repair performance, respectively. The geometric detector performs substantially worse than LLM-RobotPath in the Hard setting (50.52% vs. 93.66% BAcc). Although the standalone planner achieves a 97.00% local repair success rate, combining it with GRPO further improves the rate to 99.25% in this evaluation, demonstrating the complementary contributions of the learned policy and deterministic planner.

W2: Relation to prior works.

We acknowledge that FCL (Pan et al., 2012) supports pre-execution geometric collision checking, while CHOMP (Ratliff et al., 2009) enables collision-aware trajectory optimization. However, FCL focuses on geometric collision queries, whereas CHOMP focuses on trajectory optimization. Neither is designed to provide execution-grounded collision annotations or a unified benchmark for fine-grained path diagnosis and repair. In contrast, our work provides collision labels, first-collision localization, and verified local repairs derived from simulated executions, enabling systematic evaluation across different robot configurations, environment layouts, and execution regimes.

W3.1: Repair-Context Coverage and Target Displacement

The 97.32%, 58.15%, and 1.95% reported in Figure 5 for displacement bounds of 3m, 0.25m, and 0m represent valid repair-context coverage, rather than repair success rates. For example, when obstacles are located near the robot's original target, allowing target adjustment provides additional space for local obstacle avoidance. However, restricting the displacement to 0.25m or even 0m prevents some collision cases from forming valid repair contexts under our current repair protocol. This does not indicate model repair failures, but rather a reduction in the number of repair tasks that satisfy the constraints.

W3.2 & Q3 & Q2.1: Predicted Diagnosis and Path-Continuation Verification

The repair evaluation in Table 3 uses predicted collision diagnoses as the basis for path repair. In the local repair experiments, geometric verification primarily covers the path from the repair start to the adjusted local target. The optional continuation check further verifies the connection to the next original waypoint. In the end-to-end evaluation reported below, the system iteratively performs collision diagnosis and repair over the complete waypoint sequence. The resulting paths are then executed in simulation to evaluate whether they can be completed safely.

W3.3: End-to-End Evaluation of the Complete Pipeline

To evaluate the complete learned diagnosis-and-repair pipeline, we conduct an end-to-end simulation evaluation of LLM-RobotPath w/ PF on 520 paths under the Simple Collection, covering both the in-distribution test set and unseen scenarios. Using model-predicted collision information, the evaluation includes collision prediction and localization, local repair generation, geometric verification, fallback planner (FP), and execution of the complete repaired paths. The results are summarized below:

W4.1 & Q2.2: Clarification of GRPO’s ​FRS

The GRPO final repair success rate (FRS) reported in Table 3 corresponds to the evaluation run that achieved the highest FRS across multiple repeated evaluations of the LLM-RobotPath SFT GRPO.

W4.2 & Q4.2: Availability of Appendices

Thank you for your attention. The appendix was already included in the Supplementary Material at the time of the initial manuscript submission.

Dear Reviewer,

Thank you for your time and consideration of our submission. We are immensely grateful for your insightful reviews. As the rebuttal period is drawing to a close, please do not hesitate to reach out if you have any further questions or concerns.

Best regards,

Authors

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

This work proposes RobotPath, an Isaac Sim-based platform for pre-collision diagnosis and local repair of preplanned paths for transport robots. LLM-RobotPath is also introduced, whish is designed for numerically aware collision diagnosis and verifier-guided repair. Overall, this paper presents fairly comprehensive experiments, but some of its key arguments need further clarification.

Critical strengths
  1. This work addresses a practical issue in robot navigation: even when individual waypoints lie in free space, the robot may still collide with obstacles during motion and turning because of its finite footprint and continuous motion between waypoints. This is a meaningful motivation for checking planned paths before execution.

  2. This work provides diverse robot and environment configurations, structured collision annotations, and verified repair supervision.

Critical weaknesses
  1. The authors argue that existing work primarily addresses post-collision recovery, local replanning, or trajectory-tracking correction rather than pre-collision diagnosis and repair of planned paths. However, in my opinion, collision avoidance is commonly treated as a constraint in robot motion planning, and collision checking along planned motions is an established practice. Although the paper identifies a meaningful practical failure mode, it does not explain why existing geometric methods are inadequate.

  2. This paper uses LLM to reason over path information, robot geometry, and environmental constraints. However, the necessity of using LLM is not sufficiently justified. The reported 99.09% safe-arrival rate comes from independent execution of geometrically verified expert-generated repairs, rather than end-to-end evaluation of LLM-generated repairs. More improtantly, the relative contributions of LLM, geometric verifier, and deterministic fallback planner remains open.

  3. One of the main contributions of this paper is a new framework. However, when I click on the provided URL, it takes me to an "Anonymization/Status" page, and the progress is stuck at 0%.

Questions for authors
  1. Collision-free motion is a common requirement in robot path planning. The authors should explain which limitations of existing path-validation and repair methods motivate a dedicated framework?

  2. The reason of integrating LLM-generated repair candidates with geometric verification and deterministic fallback should be explained. A comparison against non-LLM geometric collision-checking and local-replanning baselines under the same repair constraints and verification procedures is useful.

  3. Can the authors clarify whether the curriculum-learning data affect the validity of the unseen-robot transfer evaluation?

  4. The end-to-end repair success rate when collision information is predicted by the LLM should be reported. The current repair experiments mainly evaluate repair policies using ground-truth collision contexts.The authors should report an end-to-end simulation evaluation covering collision prediction, collision localization, repair generation, geometric verification, fallback, and repair execution.

  5. When I click on your provided URL, it takes me to an "Anonymization/Status" page, and the progress is stuck at 0%.

5 comments

We thank the reviewer for the valuable comments and address each point as follows:

W1 & Q1.1: Limitations of Existing Methods

Existing geometric methods can effectively assess the geometric feasibility of planned paths and generate collision-free trajectories. However, actual robot execution is also influenced by motion states, turning behavior, and control strategies. Geometric validation based on simplified motion assumptions may fail to capture collision risks arising during actual execution. For example, in warehouse transportation scenarios, even when all discrete waypoints are collision-free, the continuous swept footprint of a robot during translation and rotation may still intersect shelves, potentially causing payload loss or task interruption.

W2.1 & Q1.2 & Q2.1: Motivation for Using ​​LLMs

Our motivation for introducing an LLM is that collision risks for transport robots depend not only on static geometry but also on orientation changes and coupled translation and rotation. We therefore use structured prompts to encode the robot footprint, obstacle geometry, complete planned path, and path-segment features. Training on collision annotations obtained from actual Isaac Sim executions enables the model to learn the relationship between execution context and collision outcomes, supporting pre-execution collision diagnosis and local path repair. Results on the Hard Collection show that LLM-RobotPath achieves 93.66% balanced accuracy, substantially outperforming the geometric baseline (50.14%). These results provide empirical justification for introducing an LLM to capture execution-dependent collision risks that are challenging for the evaluated geometric baseline, particularly under coupled-motion conditions[1].

[1] Firoozi R, Tucker J, Tian S, et al. Foundation models in robotics: Applications, challenges, and the future[J]. The International Journal of Robotics Research, 2025, 44(5): 701-739.

Q2.2: Non-LLM geometric collision-checking and local-replanning baselines

For the non-LLM geometric collision-checking baseline, we use the swept-footprint collision-checking procedure underlying the Geometric Verifier (Sec. 2.2), together with controller-kinematic replay. This allows us to directly evaluate geometric collision checking on the diagnosis test sets. We further introduce two categories of baselines: (1) two LLMs (GPT-6-Astra and DeepSeek-flash v3); and (2) lightweight MLP- and GNN-based models. Specifically, the MLP processes flattened geometric features, while the GNN employs multi-head self-attention over a fully connected graph of geometric entities, followed by graph-level pooling to obtain a scene representation. Both models jointly predict collision occurrence, the colliding object, the first collision segment, and the contact location. The comparison results are presented below. All metrics are percentages except MCE, which is reported in meters.

ModelBAcc ↑F1 ↑ObjAcc ↑SegAcc ↑[email protected] ↑[email protected] ↑MCE ↓
Simple collection
Geometric Verifier95.6397.8699.7681.5135.0471.780.1723
MLP66.1382.3960.3423.841.954.871.0190
GNN91.1495.2898.0577.8614.3642.580.2640
GPT-6-astra93.9295.0192.7070.8053.7785.400.1097
DeepSeek-flash v395.6396.6895.6272.2658.3988.560.1085
LLM-RobotPath98.4299.0398.7890.5179.8191.480.0820
Hard collection
Geometric Verifier50.5281.5970.5348.123.617.052.2647
MLP67.6385.2732.762.660.160.315.6120
GNN59.8442.057.050.310.000.005.1180
GPT-6-astra48.1578.7664.2650.003.613.611.7938
DeepSeek-flash v367.4185.5160.3439.661.725.172.0199
LLM-RobotPath93.6694.9388.0976.1864.2668.501.0073

For the non-LLM local-replanning baseline, we introduce the Fallback Planner (FP), a standalone deterministic planning method. SFT and GRPO refer to supervised fine-tuning and subsequent GRPO optimization, respectively. Greedy (Top-1) uses a single greedily generated candidate, while Bo8 selects from eight sampled candidates. w/ FP invokes the fallback planner when no model-generated candidate passes geometric verification. The evaluation uses 411 collision cases from the Simple test set, of which 400 have valid repair contexts. All methods in this repair comparison use ground-truth collision context, identical repair constraints, and the same geometric verification procedure.

MethodLocal / 400All collisions / 411Fallback addedMedian generation/search (s)
Fallback Planner (FP)388/400 (97.00%)94.40%00.1252
SFT_Greedy289/400 (72.25%)70.32%00.9058
SFT_Bo8380/400 (95.00%)92.46%02.5720
SFT_Bo8 w/ FP396/400 (99.00%)96.35%162.5720
GRPO_Greedy375/400 (93.75%)91.24%00.8281
GRPO_Bo8385/400 (96.25%)93.67%01.1328
GRPO_Bo8 w/ FP397/400 (99.25%)96.59%121.1328

The two tables report collision-diagnosis and local-repair performance, respectively. The geometric detector performs substantially worse than LLM-RobotPath in the Hard setting (50.52% vs. 93.66% BAcc). Although the standalone planner achieves a 97.00% local repair success rate, combining it with GRPO further improves the rate to 99.25% in this evaluation, demonstrating the complementary contributions of the learned policy and deterministic planner.

W2.3: Contributions of Individual Components

The LLM generates collision diagnoses and candidate repair solutions, with its repair capability further optimized through GRPO training. The geometric verifier evaluates the feasibility of these candidates and provides reward feedback for GRPO training. Specifically, GRPO_Bo8 (LLM-based) generates eight candidate repairs, which are individually checked by the geometric verifier. If all candidates fail verification, the deterministic Fallback Planner (FP) is invoked for additional search. As shown in the table below, under the same collision contexts, both GRPO_Bo8 and FP successfully repair 376 cases, while GRPO_Bo8 and FP uniquely succeed in 9 and 12 cases, respectively, demonstrating their complementary capabilities. Incorporating FP improves the overall repair success rate from 96.25% to 99.25%, validating the effectiveness of the cascaded repair mechanism.

TypesNumber of cases
Both GRPO_Bo8 and FP succeed376
Only GRPO_Bo8 succeeds9
OnlyFPsucceeds12
Both fail, including unavailable repair contexts14

Q3: Curriculum learning and unseen-robot transfer

The curriculum data is generated from sources restricted to the training split sources, excluding the held-out test robots' geometries and execution records. The held-out footprints are also excluded from subsequent task training and curriculum replay. Therefore, the reported unseen-robot evaluation does not expose the model to those test footprints through the curriculum. We have made the source restrictions explicit in the experimental description of the revised manuscript.

Q4 & W2.2: End-to-End Evaluation of the Complete Pipeline

We conduct an end-to-end simulation evaluation of LLM-RobotPath w/ PF on 520 paths under the Simple Collection, covering both the in-distribution test set and unseen scenarios. Using model-predicted collision information, the evaluation includes collision prediction and localization, local repair generation, geometric verification, planner fallback (PF), and execution of the complete repaired paths. The results are summarized below:

MethodSafe complete-path executions /520Unsuccessful pathsCollision after repairUnable to repair
Original Paths205 (39.42%)315 (60.58%)——
Geometry & PF222 (42.69%)298 (57.31%)6 (1.15%)224
LLM-RobotPath295 (56.73%)225 (43.27%)18 (3.46%)125
LLM-RobotPath w/ PF297 (57.12%)223 (42.88%)22 (4.23%)52
Reviewer_cCTcOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents RobotPath, a platform built on Isaac Sim for warehouse transport robots. The platform generates planned waypoint paths with injected collision causes such as lateral offsets, corner cutting, docking over travel and wall drift, executes them under a Simple regime with turn then translate control and a Hard regime with coupled translation and rotation, and records first collision annotations. It also provides an Expert Repair Generator that combines direct connections, analytic detours and Hybrid A* search, together with a Geometric Verifier that performs swept footprint checking. The accompanying RobotPath-Bench contains 20,399 records spanning 5 robots and 7 layouts. Results show high classification accuracy on in distribution data, degraded localisation in the Hard regime and under configuration transfer, and roughly 99% final repair success once a planner fallback is included.

Critical strengths
  • A configurable simulation pipeline with execution-derived collision labels and verified repair supervision is solid engineering; the 99.09% verifier-vs-execution agreement is a reasonable sanity check.

  • The hierarchical protocol (classification → attribute → localisation; constructibility → success → quality) and the explicit separation of model success from fallback success (Top-1/Bo8 vs FAR/FRS) are well designed.

  • Configuration-transfer splits are a good addition, and the paper is candid that transfer performance collapses.

Critical weaknesses
  • No baselines of any kind. There is no classical collision checker, no lightweight learned model (MLP/GNN on the same geometry), no other LLM, and no planner-only repair baseline.

  • Dataset realism and evaluation scale. Collisions are synthesised by scripted perturbations of otherwise valid paths rather than arising from planner or localisation failures; scenes are static shelves only.

Questions for authors
  • What accuracy does the platform’s own Geometric Verifier (or any swept-footprint check) obtain on the diagnosis test sets, and why would a practitioner prefer the LLM?

  • How does the fallback planner alone perform on Top-1/FRS and on the repair-quality metrics, and how were the separately selected GRPO numbers chosen?

3 comments

We thank the reviewer for the valuable comments and address each point as follows:

W1.1: More baselines and accuracy of Geometric Verifier.

We use the swept-footprint collision-checking procedure underlying the Geometric Verifier (Sec. 2.2), together with controller-kinematic replay, as our non-LLM collision-diagnosis baseline. This allows us to directly evaluate geometric collision checking on the diagnosis test sets. We further introduce two categories of baselines: (1) two LLMs (GPT-6-Astra and DeepSeek-flash v3); and (2) lightweight MLP- and GNN-based models. Specifically, the MLP processes flattened geometric features, while the GNN employs multi-head self-attention over a fully connected graph of geometric entities, followed by graph-level pooling to obtain a scene representation. Both models jointly predict collision occurrence, the colliding object, the first collision segment, and the contact location. The comparison results are presented in the following table. All metrics are percentages except MCE, which is reported in meters.

ModelBAcc ↑F1 ↑ObjAcc ↑SegAcc ↑[email protected] ↑[email protected] ↑MCE ↓
Simple collection
Geometric Verifier95.6397.8699.7681.5135.0471.780.1723
MLP66.1382.3960.3423.841.954.871.0190
GNN91.1495.2898.0577.8614.3642.580.2640
GPT-6-astra93.9295.0192.7070.8053.7785.400.1097
DeepSeek-flash v395.6396.6895.6272.2658.3988.560.1085
LLM-RobotPath98.4299.0398.7890.5179.8191.480.0820
Hard collection
Geometric Verifier50.5281.5970.5348.123.617.052.2647
MLP67.6385.2732.762.660.160.315.6120
GNN59.8442.057.050.310.000.005.1180
GPT-6-astra48.1578.7664.2650.003.613.611.7938
DeepSeek-flash v367.4185.5160.3439.661.725.172.0199
LLM-RobotPath93.6694.9388.0976.1864.2668.501.0073

W1.2: Motivation for Using LLMs.

Our motivation for introducing an LLM is that collision risks for transport robots depend not only on static geometry but also on orientation changes and coupled translation and rotation. We therefore use structured prompts to encode the robot footprint, obstacle geometry, complete planned path, and path-segment features. Training on collision annotations obtained from actual Isaac Sim executions enables the model to learn the relationship between execution context and collision outcomes, supporting pre-execution collision diagnosis and local path repair. Results on the Hard collection show that LLM-RobotPath achieves 93.66% balanced accuracy, substantially outperforming the geometric baseline (50.52%). These results provide empirical justification for introducing an LLM to capture execution-dependent collision risks that are challenging for the evaluated geometric baseline, particularly under coupled-motion conditions [1].

[1] Firoozi R, Tucker J, Tian S, et al. Foundation models in robotics: Applications, challenges, and the future[J]. The International Journal of Robotics Research, 2025, 44(5): 701–739.

W2.1: Dataset realism and evaluation scale.

RobotPath-Bench contains 20,399 path-execution records across 5 robot configurations, 7 shelf layouts, and 2 execution regimes, including unseen robot and layout settings to evaluate generalization. Static shelves are typical obstacles in warehouse environments, where collision risks primarily arise from robot motion relative to shelves rather than shelf movement. We construct diverse collision scenarios by varying robot geometry, shelf layouts, execution regimes, and planned paths. Controlled path generation covers representative risks such as shelf-corner clipping, lateral offsets, docking overshoot, and wall-adjacent drift, enabling systematic evaluation across different configurations and motion conditions.

W2.2: Execution-grounded collision annotations. ​

We clarify that the intended collision scenario does not determine the ground-truth label. All paths are executed in Isaac Sim, where collision occurrence, the first colliding object and segment, and contact position are determined from execution logs and contact events. Thus, a path designed to induce a collision may execute safely, and vice versa. This ensures that collision annotations reflect actual simulated execution outcomes rather than predefined collision labels.

Q1.2 & Q2.1: Detailed ​fallback​ planner.

To distinguish the contributions of the deterministic planner and learned repair policies, we conduct a controlled comparison under identical repair constraints. The evaluation uses 411 collision cases from the Simple test set, of which 400 have valid repair contexts. All methods use ground-truth collision annotations, the same geometric verification procedure, and a maximum target displacement of 3m, isolating repair performance from diagnosis errors. ​Fallback​​ Planner (​FP) denotes the standalone deterministic planner. SFT and GRPO refer to supervised fine-tuning and subsequent GRPO optimization, respectively. Greedy (Top-1) uses a single generated candidate, while Bo8 selects from 8 sampled candidates. w/ FP invokes the fallback planner when no model-generated candidate passes geometric verification.

MethodLocal / 400All collisions / 411Fallback addedMedian generation/search (s)
Fallback Planner (FP)388/400 (97.00%)94.40%00.1252
SFT_Greedy289/400 (72.25%)70.32%00.9058
SFT_Bo8380/400 (95.00%)92.46%02.5720
SFT_Bo8 w/ FP396/400 (99.00%)96.35%162.5720
GRPO_Greedy375/400 (93.75%)91.24%00.8281
GRPO_Bo8385/400 (96.25%)93.67%01.1328
GRPO_Bo8 w/ FP397/400 (99.25%)96.59%121.1328

Q2.2: Clarification of GRPO’s ​FRS.

The GRPO final repair success rate (FRS) reported in Table 3 corresponds to the evaluation run that achieved the highest FRS across multiple repeated evaluations of the LLM-RobotPath SFT GRPO.

[1] Firoozi R, Tucker J, Tian S, et al. Foundation models in robotics: Applications, challenges, and the future[J]. The International Journal of Robotics Research, 2025, 44(5): 701-739.

Dear Reviewer,

Thank you for your time and consideration of our submission. We are immensely grateful for your insightful reviews. As the rebuttal period is drawing to a close, please do not hesitate to reach out if you have any further questions or concerns.

Best regards,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.