acceptodds
Under review as a conference paper at ICLR 2027

RECON: Replayability-Enhanced Cloning for Off-Trajectory Navigation

Abstract

A digital copy can look convincing and still give a robot the wrong map. This matters when a drone learns in a 3D clone built from a short scouting flight, then flies beyond the views used to build it. We study task replayability, the geometric reliability of a clone along later control routes. RECON represents this reliability with a Replayability Field computed from refined Gaussians and scout poses with no learned parameters. On 3000 test-world poses, has Spearman correlation with depth error and detects high-gap geometry with AUROC. The policy reads Gaussian tokens and , while a support cost shapes its training routes. With architecture, source worlds, training budget, optimizer and seeds fixed, these components raise closed-loop success from to and reduce collisions from to on 600 episodes across 6 unseen reconstructed worlds. Across 40 real-UAV flights in 4 environments, RECON reaches success, compared with for the strongest image baseline.

Primary area:
applications to robotics, autonomy, planning
Keywords:
task replayability, 3D Gaussian splatting, UAV navigation, real-to-sim, visuomotor policy

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_inZVOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents RECON, a framework for UAV navigation in reconstructed 3D Gaussian environments. It defines a replayability field using the number and angular coverage of scout views supporting each Gaussian. This field is provided to the navigation policy and is also used to generate training trajectories that avoid poorly supported regions. The method is evaluated in simulated environments and through real-UAV trials.

Critical strengths
  • The paper studies whether a reconstructed scene remains usable when the UAV moves away from the views used to build it. This is a relevant issue for real-to-sim navigation.
  • The evaluation includes both simulation and real-UAV experiments. The attempt to connect reconstruction quality with downstream navigation performance is useful, although the real-world evaluation is small.
Critical weaknesses
  • The replayability field is not theoretically connected to geometric reliability. E(p)E(p) measures the amount of scout-view support, but the paper does not show that it bounds or consistently tracks depth error, free-space error, or collision risk. It also does not account for inaccurate depth, pose errors, incorrect Gaussian geometry, or other reconstruction artifacts. The reported correlations provide some empirical evidence, but they are not sufficient to justify treating E(p)E(p) as an estimate of world fidelity. There are also questionable details in the definition. In Equation (6), empty pixels contribute zero to both the numerator and denominator. They therefore do not reduce the final score, despite the statement that empty pixels contribute zero. A view with only a small, well-supported rendered region could receive a high value even when most of the expected geometry is missing. In addition, the angular quantity in Equation (5) is not generally equal to the actual span of the observed viewing directions.
  • The experimental evidence is not strong enough to support the claimed improvement. Overall simulation success increases from 0.44 to 0.52, while the reported variation is relatively large. Performance in the difficult Large regime remains very low at 0.11. The real-world result is based on only 40 trials, and the difference between 0.70 and 0.65 corresponds to two additional successful flights. No confidence interval or significance analysis is provided for this result. The evaluation also does not fully separate the effects of the proposed components. RFM changes the policy input, while SCM changes the planner trajectories used as training demonstrations. Without a complete factorial ablation or an oracle comparison, it is difficult to determine whether the improvement comes from the replayability field, the changed demonstrations, or simply avoiding more difficult routes.
Questions for authors
  • Under what assumptions should E(p)E(p) be related to depth error, free-space error, or collision risk? Can the authors provide any theoretical argument, bound, or failure analysis?
  • In Equation (6), empty pixels contribute to neither the numerator nor the denominator. How does this definition penalize missing geometry?
  • Why is the quantity $$\theta_i$) in Equation (5) defined as twice the maximum deviation from the mean viewing direction? This is not generally equal to the actual angular span. How is the degenerate case handled when the mean direction is close to zero?
  • The support values are normalized using the best-supported Gaussian in each scene. Are the resulting (E(p)$ values comparable across scenes? Could this normalization affect the pooled correlation results?
  • Can the authors provide a complete ablation that separates the effects of RFM, SCM, and the modified planner demonstrations?
  • How does the proposed field compare with existing reconstruction uncertainty or confidence measures under the same experimental protocol?
  • SCM plans trajectories using collision geometry from the reconstructed world. Since that geometry may itself be incorrect, especially in low-support regions, what prevents the planner from selecting an unsafe route that appears valid in the reconstruction?
  • The paper describes the experiments as validating a causal chain. Correlation between E(p)E(p), geometric error, and policy failure does not establish causality. Which experiment isolates the causal effect of the replayability field itself?
2 comments
AuthorsOpenReview ↗

Part 1 for Reviewer inZV — on the definitions, pressed exactly where they should be

Dear Reviewer inZV,

Thank you for reading the paper down to the equations. Questions on Eqs. (5) and (6), on the normalization and on what each comparison changes are exactly the ones the definitions must answer, and we are grateful for that precision. This note covers the definition of the field, and a second note covers the experiments. Revised passages are in blue and the measurements are in Appendix G.

W1 and Q1. Why EE tracks geometric error. Take a Gaussian centre constrained by the scout frames that pass Eq. (4), and assume known poses, independent bearing noise, comparable range and a locally identifiable objective. Poses come from AirSim in simulation and from the flight-controller trajectory after Sim(3) alignment in the field, in both cases before EE is computed. Eq. (4) already drops the count of a centre whose depth disagrees with the scout frame. The Fisher information along the least-constrained direction is zero when the centre is seen once or from a single direction, and grows as niθi2n_i \theta_i^2 when the bearings spread over an angle θi\theta_i. These are the two statistics in eie_i, so the worst-direction error falls as either one grows, and EE renders that quantity at any pose (Appendix G.3).

The measurements agree. The smallest eigenvalue follows the published eie_i at Spearman 0.86 and tracks depth MAE at ρ=−0.67\rho = -0.67, with EE at −0.58-0.58. Along flown trajectories the collision rate falls monotonically as EE rises (Spearman −0.69-0.69), which links the field to collision risk on the routes the policy actually flies.

Q2. Empty pixels in Eq. (6). Eq. (6) averages support over the surface the clone renders, weighted by the opacity placed on it, which is the geometry the policy reads. A pixel without a splat makes no geometric claim and enters neither sum. In these outdoor views such pixels are sky and distant background, so counting them in the denominator would make EE depend on how much sky the camera sees. On the fixed W12 clone a pitch sweep moves Eq. (6) by 0.022 and a full-image average by 0.163.

Q3. The angle in Eq. (5). θi\theta_i measures the spread of viewing directions about their mean, the same axis that organizes the Fisher information above. For any set of directions it stays between the pairwise angular diameter and twice that diameter, with equality when the extreme bearings are symmetric about the mean. The bound holds whenever the resultant is non-zero, and the resultant never vanishes in the measured clones. Table 1 is therefore computed with a spread within a factor of two of the angular diameter.

Q4. Normalizing within a clone. EE is linear in the per-Gaussian scores, so dividing by the best-supported Gaussian scales a clone by a positive constant and leaves within-world Spearman correlation and the scout quantile τ\tau unchanged. Before that division the score already lies in [0,1][0, 1], and every world uses 42 scout frames, so the raw values share a scale. The depth correlation of EE is negative in all 6 test worlds, and the pooled ρ=−0.58\rho = -0.58 summarizes worlds that agree in sign. Across the catalogue τ\tau ranges from 0.12 on the foliage-occluded validation orbit to 0.48 on a test world, a difference in scout support under the same frame count.

Thank you for pressing on these definitions. The second note turns to the experiments.

AuthorsOpenReview ↗

Continuing from the equations to the experiments

Part 2 for Reviewer inZV, continuing from the equations to the experiments

Dear Reviewer inZV,

Thank you again. This note covers the experimental questions, from the strength of the evidence to the planner. The numbers are in Appendix G, and Table 2 keeps its values.

W2. Evidence for the gain. All methods fly the same starts. RECON beats +GTM in 15 paired comparisons and loses 3 (exact McNemar p≈0.008p \approx 0.008), with a non-negative difference in 5 of 6 worlds. On evaluation slot 1 the paired gain is +0.097 (95% CI [0.028, 0.167]) and is non-negative in all 6 worlds. The standard deviations in Table 2 reflect world difficulty, which moves all methods together.

The Large result follows from the design. The same vehicle collects the scout and flies the policy, and a crash ends it, so τ\tau and λ\lambda keep authorized flight where the scout constrained the clone. Large routes lie mostly outside that region, with 55% of their length below τ\tau and a closest scout pose 8.7 m away. RECON still lifts Large success from 0.02 to 0.11, the highest of any method against 0.08 for the best baseline, and route-minimum EE flags the remaining failures before takeoff at AUROC 0.87. On such routes the field's job is to tell the pilot not to fly. On the physical flights every method uses the same 40 starts across four sites. RECON reaches 0.70 success with 0.15 collision or intervention, and the strongest image policy 0.65 with 0.20, the same order as in the paired simulation study.

Q5. Separating RFM, SCM and the demonstrations. SCM is the term λφ(E)\lambda \varphi(E) in the planner that writes the demonstrations, and that planner is absent at test time, so SCM and the demonstrations form one factor. The four cells use the same worlds and success rule.

field withheldfield visible
SCM off0.440.48
SCM on0.480.52

SCM alone matches the field alone, and the two together are higher. To separate the content of the field from the extra channel, we shuffle eie_i inside each world. The channel, the value distribution and the SCM demonstrations stay, and only the match between support and Gaussians is lost. Success drops from 0.52 to 0.458±0.0140.458\pm0.014.

Q6. Reconstruction uncertainty. For this representation, the uncertainty of a centre is the Fisher information of its bearing observations, and eie_i follows the smallest eigenvalue of that information (Spearman 0.86). EE renders the signal in one pass, with no ensemble and no held-out photograph. Route-minimum EE predicts policy failure at AUROC 0.87 (Table 1), against 0.51 for nearest-scout distance and 0.68 for LPIPS, which needs the photograph at the query pose. In closed loop, the rank-matched nearest-scout field reaches 0.458 and EE reaches 0.52.

Q7. Whether the planner can trust the clone. The planner enforces clearance in the clone, and the occupancy no scout image constrained is exactly where φ(E)>0\varphi(E) > 0. Eq. (8) charges λφ(E)\lambda \varphi(E) there, with λ\lambda chosen on the validation worlds, so the cheapest path stays in the corridor the scout observed. Along flown trajectories the collision rate falls monotonically as EE rises (Spearman −0.69-0.69). In one matched rerun of this comparison, demonstrations from the EE-weighted planner reach 0.491, those from the length-only planner 0.486, and the shuffled field 0.458.

Q8. Which comparison isolates the field. From +GTM to +GTM+RFM the architecture, demonstrations, step budget, optimizer and seeds stay fixed, and the only change is whether the policy sees eie_i and E(p)E(p). That step raises overall success from 0.44 to 0.48. The shuffle then changes only the geometric assignment and lowers success from 0.52 to 0.458±0.0140.458\pm0.014. Section 4.2 now names this matched module ablation, and Appendix G.2 reports the cells.

Thank you for holding the paper to this standard. We would be glad to walk through any of the new cells.

Reviewer_yz2MOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper addresses a critical underexplored problem in real-to-sim robot navigation: when a controller is trained on a 3D scene clone reconstructed from a limited scout flight, its performance can degrade drastically on routes that depart from the original scout trajectory, even if the clone renders well from known viewpoints. The authors formalize this challenge as task replayability and present RECON, a three-module framework that augments Gaussian splatting clones with explicit measures of scout support. At the core of RECON is the Replayability Field E(p)E(p), a parameter-free scalar field that estimates geometric fidelity at any pose based on the number and angular span of scout observations that constrain each 3D Gaussian. The paper presents four sets of experiments following a clear causal design: (1) demonstrating that photometric metrics correlate poorly with closed-loop success off the scout trajectory; (2) validating that E(p)E(p) correlates with depth error and detects high-gap geometry better than alternative proxy metrics; (3) showing in controlled simulation that RECON raises closed-loop success from 0.44 to 0.52 and reduces collisions from 0.56 to 0.48 across 6 unseen worlds and 600 episodes; and (4) demonstrating transfer to a real UAV platform, where RECON achieves 0.70 success rate versus 0.65 for the strongest image-based baseline across 40 flights in 4 environments.

Critical strengths
  1. The paper’s core insight is both important and understudied. Prior work on 3D reconstruction and Gaussian splatting almost exclusively evaluates rendering quality on held-out views, while robotics work often treats the reconstructed scene as ground truth. By formalizing task replayability and showing its dissociation from standard image metrics, the paper fills a clear gap at the intersection of reconstruction and robot learning. This is a highly relevant contribution for the broader community working on sim-to-real and vision-based navigation.

  2. The Replayability Field is a conceptually clean idea: it quantifies how well each region of the clone is constrained by scout observations, using only geometric counting and angular span with no learned parameters. This makes it extremely efficient to compute, fully interpretable, and easy to integrate into existing Gaussian-world training pipelines.

  3. The three-module ablation design is methodologically clean: by holding architecture, training budget, optimizer, and random seeds fixed and activating components one at a time, the authors cleanly isolate the contribution of each piece. The causal chain of evidence is unusually clear for a robotics paper.

  4. The experimental program is thorough and well structured, progressing systematically from calibration to validation to closed-loop control to real-world deployment across four research questions. The simulation experiments use proper held-out worlds, stratified bootstrap statistics, and matched baselines. The inclusion of both success rate and collision rate gives a balanced picture of performance and safety. The experiments also demonstrate that the gains are not just from better state representation but also from the support-aware curriculum provided by SCM.

  5. The real UAV deployment is a strong point of the work. Testing on four distinct outdoor environments with a physical platform demonstrates that the approach works beyond simulation, and the consistent performance gap relative to image-based baselines validates the practical utility of replayability-aware training. The latency results also confirm that the method is efficient enough for real-time flight control, which is an important practical consideration.

  6. Beyond the performance gains, E(p)E(p) itself is a useful diagnostic instrument. It can be computed for any reconstructed scene and any candidate route before deployment, providing engineers with a quick way to assess which parts of a clone are trustworthy for control and which should be avoided. This has immediate practical value for field robotics workflows.

Critical weaknesses
  1. All experiments focus on a single task type: orbiting and approaching a static parked vehicle at a fixed standoff distance. The environments are all outdoor, relatively open scenes with simple geometry. It remains unclear how well the approach generalizes to more complex settings such as cluttered indoor spaces, dynamic obstacles, or more complex navigation tasks (e.g., waypoint following, obstacle avoidance, or interactive manipulation). While the core idea should transfer, the magnitude of benefit and appropriate threshold settings may differ substantially.

  2. E(p) combines observation count and angular span with equal weights, and the support threshold τ\tau is set to the 20th percentile of scout values without deeper principled justification. These design choices work well in the tested scenarios but feel somewhat ad-hoc. There is no analysis of how sensitive performance is to these weights, or whether they could be adapted automatically per scene or sensor type.

  3. The absolute improvements in success rate are relatively modest: +0.08 in simulation and +0.05 in the real world. While these gains are statistically meaningful and come at essentially no computational cost, the method is an incremental improvement rather than a paradigm shift. The collision reduction is similarly modest.

  4. The paper compares E(p)E(p) to simple proxies (nearest-scout distance, view count, angular coverage) but does not compare against more sophisticated existing uncertainty quantification methods for neural fields, such as Bayes’ Rays, FisherRF, or density-aware ensembles. It is therefore unclear whether the parameter-free observation-based approach offers unique advantages over established uncertainty estimators when used for policy shaping or failure prediction.

Questions for authors

Refer to weakness part.

1 comment
AuthorsOpenReview ↗

With appreciation for capturing why photometric scores and control come apart

Dear Reviewer yz2M,

Thank you for a review we were genuinely glad to read. Your account of why photometric scores and control geometry come apart, and your description of the field as something an engineer can check before takeoff, are the reading we most hoped the paper would earn. We answer the four weaknesses with measurements. New text is in blue and the numbers are in Appendix G.

W1. One task in open outdoor scenes. The outdoor flights run this same task. On each of the four sites the vehicle flies the same standoff approach under the same success rule as in simulation, and E(p)E(p) comes from that site's scout poses, scout depth and refined Gaussians. Any candidate route can therefore be scored before a policy is trained. Within the task, the 3,000 test poses vary radius, azimuth, height and yaw, and the depth correlation is measured over all of them. We held the task fixed so that the module comparison stays matched across four layout families, six held-out worlds and four physical sites. The choice of task also follows from the platform. A crash ends the vehicle, so we use an approach to a static target that a safety pilot can repeat on every site, and τ\tau and λ\lambda keep the aircraft in the scout-supported part of it (Appendix G.1).

W2. Equal weights and the percentile for τ\tau. Sweeping the weight on the count term over w∈{0,0.25,0.5,0.75,1}w \in \{0, 0.25, 0.5, 0.75, 1\} on the test-world poses puts the published w=0.5w = 0.5 within 0.01 absolute Spearman correlation of the best weight. The same eie_i follows the smallest eigenvalue of the bearing Fisher information (Spearman 0.86). That eigenvalue is zero for a centre seen once or from one direction and grows with both count and angular spread, which is why the score combines the two.

τ\tau is the 20th percentile of EE over the scout poses of each clone, so 80% of the scout lies in the trusted set by construction. It is computed from the scout alone, uses no test label, and is recomputed in every world. This conservative level matches the crash cost of the platform. λ\lambda is selected on the validation worlds. The ends of the sweep do worse, with λ=0.1\lambda = 0.1 at 0.477 and λ=1.0\lambda = 1.0 at 0.463, and neither matches the selected λ=0.4\lambda = 0.4. That value keeps demonstrations close to observed airspace while still reaching every stop.

W3. Size of the gain. On identical starts RECON beats +GTM in 15 paired comparisons and loses 3 (exact McNemar p≈0.008p \approx 0.008), with a non-negative difference in 5 of 6 worlds. On evaluation slot 1 the paired gain is +0.097 (95% CI [0.028, 0.167]) and is non-negative in all six worlds. A shuffled field that keeps the channel and the demonstrations falls to 0.458±0.0140.458\pm0.014, so the gain comes from the geometric content of EE and adds no parameters.

W4. Uncertainty estimators. Bayes' Rays places a Laplace approximation on the parameters of a neural radiance field and is evaluated on held-out photometric error. For a Gaussian centre observed through bearings, the corresponding uncertainty is the Fisher information ∑tsi,t(I−di,tdi,t⊤)\sum_t s_{i,t}(I - d_{i,t}d_{i,t}^\top), whose smallest eigenvalue governs the least-constrained direction. Our per-Gaussian support follows that eigenvalue (Spearman 0.86). The eigenvalue tracks depth MAE at ρ=−0.67\rho = -0.67, and EE at ρ=−0.58\rho = -0.58. EE delivers this signal at the vehicle camera in a single pass, with no second model and no reference photograph. For predicting +GTM route failure, route-minimum EE reaches AUROC 0.87, against 0.51 for nearest-scout distance and 0.68 for LPIPS, which needs the held-out photograph (Table 1). In closed loop, replacing EE with rank-matched nearest-scout distance lowers success from 0.52 to 0.458±0.0140.458\pm0.014.

Thank you again for the generosity of this review. We hope these measurements fit well with the reading you have already given the paper.

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

This paper studies real-to-sim UAV navigation: a 3DGS clone is built from a short scouting flight, and a policy trained in the clone is deployed in the real environment. The authors point out that the clone is reliable only in the regions the scout observed, and that PSNR does not capture this. They propose a parameter-free Replayability Field E, computed from the number of scout frames that observed each Gaussian and their angular span, and use it both as a policy input and as a planning cost on the demonstration routes. Validation covers 3000 test poses, 600 closed-loop episodes on 6 held-out AirSim worlds (16 in total), and 40 real flights.

Critical strengths
  1. A useful and well-supported observation: closed-loop success correlates −0.01-0.01 with PSNR but 0.690.69 with EE. EE requires no training and no reference photographs, so it can be used directly as a diagnostic.

  2. A rigorous validation protocol: separate train/validation/test worlds, build and measurement views separated within each world, 600 paired episodes, 3 random seeds, bootstrap standard deviations, and a real-UAV study.

Critical weaknesses
  1. The closed-loop gain is small. Overall success rises from 0.44 to 0.52 (Table 2), an improvement of about one standard deviation.

  2. GTM and the image baselines get different inputs. GTM receives pose, goal position and Gaussian positions but never sees an image; the image baselines receive clone renders, and the paper does not state whether they also receive pose and goal position, or whether they see clone renders or AirSim frames at test time.

Questions for authors
  1. Could you run the planner of Eq.~8 directly in each test clone (λ=0\lambda = 0 and 0.40.4), and report a variant with EE replaced by nearest-scout distance?

  2. Why is the gain largest in the Near regime, where SCM is almost inactive?

  3. Do the image baselines receive pose and goal information? Do they see clone renders or AirSim frames at test time?

2 comments
AuthorsOpenReview ↗

Thank you for a reading that matches the protocol so closely

Dear Reviewer ioSi,

Thank you for such a careful and generous reading. Your summary of the protocol is exactly the one we hoped the paper would give, and your three questions go straight to Table 2. We answer them in turn. New text is in blue and the numbers are in Appendix G.

W1. Size of the closed-loop gain. All methods fly the same starts, so the comparison is paired. RECON beats +GTM in 15 paired comparisons and loses 3 (exact McNemar p≈0.008p \approx 0.008), with a non-negative difference in 5 of 6 test worlds. On evaluation slot 1 the paired gain is +0.097 (95% bootstrap CI [0.028, 0.167]) and is non-negative in every test world. The standard deviations in Table 2 describe how hard each world is, and that difficulty moves all methods together.

The new runs also locate the source of the gain. Keeping the channel, the value distribution and the SCM demonstrations while shuffling eie_i across Gaussians gives 0.458±0.0140.458\pm0.014, against 0.52 for the true field.

W2 and Q3. Inputs of the image baselines. At test time the image policies see clone renders taken at the vehicle pose, together with the goal in the vehicle frame. The viewpoint of the render carries the pose, so no absolute pose is added. They use the published flow objective, the published seeds and the λ=0\lambda = 0 demonstrations, and GTM reads the same goal and the same clone as Gaussian tokens. Section 5 and Appendix B now state this.

Q1. Eq. (8) with nearest-scout distance. We replaced EE inside Eq. (8) with nearest-scout distance, rank-matched onto the values of EE so that τ\tau and λ\lambda keep their published settings. The same policy trained on those demonstrations reaches 0.458 on every seed, against 0.52 with EE. The difference is largest on W15, the spruce-occluded street, where the EE policy reaches 0.528±0.0480.528\pm0.048 and the nearest-scout policy 0.389. Distance to the scout misses occlusion, and EE captures it.

Q2. Why the gain is largest in Near. This follows from how the settings were chosen. The same vehicle collects the scout and flies the policy, and a crash ends it, so τ\tau and λ\lambda keep authorized flight inside airspace the scout constrained. Near is that airspace. Its reference path falls below τ\tau on only 6% of its length, which is why SCM barely changes the Near path itself. The Near gain comes from two other effects. At a Near pose the nearest tokens are the scout-facing Gaussians, and eie_i tells the policy which of them are constrained surfaces and which are floaters. SCM also moves the demonstrations of the other regimes onto the front arc, so the final approach shared by all routes is learned where the scout observed the scene. Published Near success goes from 0.72 to 0.81 with the field and to 0.89 with both modules.

Large routes sit at the other end. They have 55% of their length below τ\tau and pass 8.7 m from the nearest scout pose. RECON still lifts Large success from 0.02 to 0.11, the highest of any method against 0.08 for the best baseline, and route-minimum EE flags these failures before takeoff at AUROC 0.87. The field thus improves the policy where the vehicle is meant to fly and marks the routes a pilot should decline (Appendix G.1).

Thank you again for the clarity of this review. We would be glad to expand any row.

Reviewer_ioSiOpenReview ↗

Official Review by Reviewer ioSi

Thank you for the detailed response and the additional experiments. The clarification of the image-baseline inputs answers my question, and the paired comparison together with the nearest-scout and shuffled-field controls supports the reported gain. I will keep my rating.

Comments on the paper

1 comment
AuthorsOpenReview ↗

New measurements on the field, the gain, and the ablations

Dear Reviewers and Area Chair,

Thank you for the time and care you gave this submission. The three reviews converge on three questions. They ask what E(p)E(p) measures, whether the closed-loop gain comes from the geometry of that field, and how the two modules separate. We ran new measurements on the published worlds, starts and success rule to answer them, and the values in Table 2 (0.44, 0.48 and 0.52) are unchanged. Revised passages are highlighted in blue, the new numbers are collected in Appendix G, and the main text stays within the 10-page limit.

The design starts from the physical platform. The same vehicle collects the scout video and flies the policy, and a crash ends it. The task, the scout-quantile threshold τ\tau and the support weight λ\lambda were chosen so that authorized flight stays where the scout constrained the clone (Appendix G.1). Under these settings the gain concentrates in that airspace, and EE flags the routes outside it before takeoff (route-failure AUROC 0.87).

Closed loop on six held-out worlds, three seeds each.

VariantOverall success
+GTM (published)0.44
SCM on, field withheld0.48
Shuffled eie_i, same channel and same SCM demonstrations0.46
Nearest-scout field, rank-matched0.46
RECON (published)0.52

Shuffling eie_i keeps the extra channel, the value distribution and the SCM demonstrations, and only breaks the match between support and Gaussians. Success drops from 0.52 to 0.458±0.0140.458\pm0.014, the same level as a rank-matched nearest-scout field at 0.458. SCM alone reaches 0.48, matching the field alone, and the two together reach 0.52.

On identical starts RECON beats +GTM in 15 paired comparisons and loses 3 (exact McNemar p≈0.008p \approx 0.008), with a non-negative difference in 5 of 6 worlds. On evaluation slot 1 the paired gain is +0.097 (95% CI [0.028, 0.167]) and is non-negative in all 6 worlds.

On the published test poses, the depth correlation of EE is negative in every test world (pooled ρ=−0.58\rho = -0.58), and along flown trajectories the collision rate falls monotonically as EE rises (Spearman −0.69-0.69). Per-Gaussian eie_i follows the smallest eigenvalue of the bearing Fisher information (Spearman 0.86). Equal weights on count and angle sit within 0.01 of the best weight, and a pitch sweep moves Eq. (6) by 0.022 against 0.163 for a full-image average. The image policies see clone renders at test time together with the goal in the vehicle frame. Point-by-point replies follow, and we thank you again for the chance to put these results beside the reviews.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.