acceptodds
Withdrawn from ICLR 2027

FaB-Fly: Foresight and Backtracking for Aerial Vision-and-Language Navigation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

In aerial vision-and-language navigation, an unmanned aerial vehicle (UAV) follows natural-language instructions that specify landmarks and how to move relative to them. Recent memory-based hierarchical approaches retrieve candidate landmark locations from memory and use a planner to select successive targets. However, several locations may match the same landmark description, so the planner must assess whether flying to each candidate would satisfy the instruction. If the UAV moves on after arrival without checking whether its flight followed the instruction, an incorrect choice can lead to further navigation errors. We propose FaB-Fly, a hierarchical navigation framework that uses Foresight to select targets and Backtracking to recover from incorrect target choices. Foresight uses a video world model to generate a flight video for each candidate, based on current observations and a candidate-specific movement instruction. A multimodal large language model (MLLM) scores how well each predicted video matches the corresponding part of the navigation instruction to guide target selection. After execution, Backtracking evaluates the observed flight against the same instruction segment. If the score is low, the UAV returns to the departure point and retries with another candidate. Experiments on AerialVLN-S and AerialVLN-Fine show that FaB-Fly outperforms the strongest baselines in success rate. Further analysis shows that FaB-Fly achieves a success rate close to that of exhaustively visiting every candidate while requiring about half the computation time.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-and-Language Navigation, Multi-modal Large Language Model

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_WFjYOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposed a zero-shot hierarchical navigation system named FaB-Fly to overcome the problem of finding candidate landmarks and avoid building up navigation errors for drone visual-language navigation. FaB-Fly uses Foresight to predict and filter candidate routes, then uses Backtracking to confirm the actual flying track, and can return to retreat and try again if necessary. The experimental results show that FaB-Fly has the leading navigation success rate on AerialVLN-S and AerialVLN-Fine.

Critical strengths

This paper aims to address the key challenges of layered navigation, namely judging whether a flight mission conforms to requirements for multiple eligible candidate targets before flight execution. Therefore, it is compatible with the demands of reality.

When estimating the computational cost of using Foresight and Backtracking,the author compared them to Visit-All, for both computational time per step and SR, and found that FaB-Fly provides a favourable trade-off in cost versus performance.

Critical weaknesses

In the ablation study, removing Foresight increases the success rate from 15.1% to 16.3%, while the full model mainly achieves a lower NE. Can the authors explain this phenomenon, and why NE is prioritized over SR? Does this choice align with the main goal of the task, which is the combination of Foresight and Backtracking?

The paper limits the number of backtracks for each instruction segment to just one. Can the authors explain the reasoning behind this design? If the second candidate location is still wrong, the system moves directly to the next instruction segment. How do you detect and prevent errors from accumulating in subsequent navigation? Also, would allowing multiple adaptive backtracks further improve performance?

Table 5 shows that the default closed-source Gemini 3.1 Pro doesn't outperform the open-source Qwen 3.8 27B and Gemma 4 26B A4B in SR, nDTW, or NE. But the paper doesn't explain why it was chosen as the default evaluator, and there's also no fair comparison of the models' capabilities, inference costs, and API overhead.

Questions for authors

In the ablation study, removing Foresight increases SR from 15.1% to 16.3%, whereas the full model mainly improves NE. Could the authors explain this trade-off and clarify why the reduction in NE is considered more important than the decrease in SR?

What motivates limiting Backtracking to a single retry per instruction segment? If the alternative candidate is also incorrect, how does the system detect and prevent the resulting error from propagating to subsequent instruction segments? Have the authors evaluated multiple or adaptive backtracking attempts?

Table 5 shows that the default closed-source Gemini 3.1 Pro does not outperform the open-weight Qwen 3.8 27B or Gemma 4 26B A4B on SR, nDTW, or NE. Why was Gemini selected as the default judge? Could the authors provide a controlled comparison of these models in terms of accuracy, inference latency, computational or API cost, and performance stability?

Reviewer_cy5JOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FaB-Fly, a zero-shot hierarchical framework for aerial vision-and-language navigation (Aerial VLN) that addresses target selection under landmark ambiguity and post-execution error correction. Its premise is that reaching a landmark that matches the appearance description does not mean the preceding flight satisfied the motion constraints in the instruction. For example, "approach a white building" and "cross the river and then approach a white building" are different tasks.

The framework decomposes the instruction into segments, retrieves candidate locations for each segment from a landmark memory built through a prior survey, and narrows them with a geometric heuristic. Foresight then generates a predicted flight video for each candidate from the current observation and a candidate-specific movement prompt, and a multimodal large language model (MLLM) judges how well each video matches the current instruction segment to select the target. After the actual flight, Backtracking evaluates the observed video; if its score falls below a threshold, the UAV returns to its departure pose and switches to another candidate, with at most one retry per segment.

Critical strengths

1. Clear motivation, addressing meaningful ambiguity and recovery problems in UAV navigation

The paper clearly distinguishes between "reaching a landmark of the correct category" and "reaching the landmark in the way the instruction requires." For instructions involving crossing, detouring, or turning, the image at arrival may not expose process errors, so verifying the flight history is reasonable. Connecting such process constraints with candidate landmark selection is a direction worth studying.

2. Well-motivated candidate-specific conditioning

The authors construct the movement prompt u(g)u(g) from each candidate's relative direction, distance, and landmark description, rather than feeding the instruction segment sis_i under verification directly into the generator. The paper clearly explains this choice: if the generator were directly asked to depict "crossing the river" and the judge then checked whether the river was crossed, all candidates could receive uninformative positive predictions. Candidate-specific prompts help reduce the risk that predictions simply cater to the verification target.

Critical weaknesses

1. The benefit of combining the components is concentrated in NE, the basis for choosing the default configuration is inconsistent, and uncertainty estimates are missing

Table 3 reports the following:

MethodSR ↑\uparrownDTW ↑\uparrowNE ↓\downarrow
w/o Both14.310.4116.3
w/o Foresight (Backtracking only)16.39.4111.4
w/o Backtracking (Foresight only)15.010.5111.3
Full FaB-Fly15.19.197.6

These results support the positive effect of Foresight used alone, as well as the value of the combination in reducing the average navigation error. However, adding Foresight on top of Backtracking lowers SR from 16.3 to 15.1 and nDTW from 9.4 to 9.1, while NE improves from 111.4 to 97.6.

Table 4 shows that this pattern is not specific to one configuration. All four judge-input combinations have lower NE (89.8–102.6) than both single-component variants (111.3 and 111.4), but all of them have lower SR than w/o Foresight (16.3), and three of the four also fall below w/o Backtracking (15.0). When only one judge uses video, SR drops to 13.0 and 11.7, below not only the image-only configuration (14.4) but also w/o Both (14.3). The benefit of the combination therefore appears consistently in NE, but not in SR or trajectory quality.

Moreover, the default configuration is not the best on any single metric across the paper's ablations. The highest SR is achieved by w/o Foresight (16.3, Table 3) and the lowest NE by "video Foresight + image Backtracking" (89.8, Table 4); in Table 5, using Qwen 3.8 27B as the judge yields a higher SR (15.7), and using Gemma 4 26B A4B yields better NE (92.3) and nDTW (10.2). When discussing Table 3, the authors use NE to argue that the combination is better than w/o Foresight, but when discussing Table 4, they use SR to argue that using videos for both judges is better than using video for only one, so the metric underlying the argument switches between the two analyses. I suggest stating the primary metric and the criterion for selecting the default configuration explicitly.

The claims should be qualified accordingly. The introduction limits the complementarity of the two components to reducing navigation errors, but the contribution list and the conclusion describe them as complementary without qualification, and the discussion of Table 4 interprets the SR result of using videos for both judges as consistent with this complementarity.

Finally, the manuscript does not report the number of repeated runs, random seeds, confidence intervals, or the number of episodes in each evaluation split. Both video generation and MLLM inference are stochastic, while the SR gain of the full system is only 15.1−14.3=0.815.1 - 14.3 = 0.8 percentage points over w/o Both and 15.1−15.0=0.115.1 - 15.0 = 0.1 points over Foresight only.

NE is also sensitive to the configuration: changing only the judge model makes NE vary between 92.3 and 107.1, a range (107.1−92.3=14.8107.1 - 92.3 = 14.8) comparable to the NE improvement of the combination over w/o Foresight (111.4−97.6=13.8111.4 - 97.6 = 13.8). Without uncertainty estimates, it is difficult to judge whether these differences are stable.

2. The improvement over external baselines comes mainly from the base framework, and the comparison conditions on AerialVLN-Fine are asymmetric

All analyses in Section 4.3, including Table 3, are conducted on AerialVLN-Fine. The SR of w/o Both already reaches 14.3, i.e., 14.3−8.0=6.314.3 - 8.0 = 6.3 points above FineCog-Nav, whereas the full method adds only 15.1−14.3=0.815.1 - 14.3 = 0.8 points over w/o Both. Hence, the 15.1−8.0=7.115.1 - 8.0 = 7.1-point improvement over FineCog-Nav comes mainly from the base framework (landmark memory, instruction grounding, geometric candidate narrowing, and low-level control) rather than from Foresight and Backtracking.

This gap is also accompanied by asymmetric comparison conditions. Page 7 states that, among the zero-shot methods, CityNavAgent, LookasideVLN, and FaB-Fly use prior environmental data, yet neither of the former two appears in Table 2; moreover, FaB-Fly's landmark memory on AerialVLN-Fine is built from images captured across each entire scene. FaB-Fly is therefore the only method in Table 2 that uses a prior survey, and its improvement over FineCog-Nav conflates prior information with the contribution of the method itself.

There are no ablations on AerialVLN-S, yet page 7 attributes the improvement over LookasideVLN to evaluating flights before and after execution. I suggest adding w/o Both on AerialVLN-S and including at least one baseline on AerialVLN-Fine that also uses environmental priors, or at least marking the priors used by each method in the tables.

The sources of the baseline numbers should also be stated: which results are cited directly and which were re-run, and which base models, environment memories, and action budgets were actually used. Differences in base models likewise affect the attribution of performance differences.

3. Direct mechanistic validation is missing for whether the signal from predicted videos reflects actual execution outcomes

The authors recognize that predictions for distant or unobserved regions may be inaccurate and use Backtracking to handle prediction errors; by design, the generated videos serve only to compare candidates and need not precisely reproduce the actual flight path (p. 2, lines 83–103; p. 3, lines 150–155). These are reasonable design choices. The remaining question is whether the key semantic information that distinguishes candidates in the generated videos can reliably predict how well the actual flight satisfies the instruction.

According to Eq. (1), V^(g)∼Wθ(⋅∣o,u(g))\hat{V}(g) \sim W_\theta(\cdot \mid o, u(g)), the video model is conditioned only on the current observation oo and the candidate-specific movement prompt u(g)u(g), whereas the actual low-level controller additionally uses the current instruction segment sis_i (p. 5, §3.2; p. 6, Eqs. (1)–(2)). The same target can be reached via different paths, so relations such as "crossing" or "detouring" depicted in the video may not match the path the controller actually takes. Imprecise predictions can still provide a useful ranking signal, but the current support comes mainly from downstream performance.

The "image Foresight" in Table 4 scores candidates from the current observation and the candidate's location information, and can serve as a control for testing whether videos provide additional information. However, the effect of video relative to images reverses depending on the Backtracking input: with image-based Backtracking, SR drops from 14.4 to 11.7; with video-based Backtracking, it rises from 13.0 to 15.1.

The explanation of Figure 3(a) also needs clarification. The movement prompt for the second candidate specifies a 10 m flight ("fly 10m toward the zebra crossing"), which, according to Section 3.3, is exactly that candidate's distance from the UAV; yet the predicted video shows the UAV not reaching the crosswalk, and the text explains this as a 10 m flight being unable to reach the second candidate.

If u(g)u(g) encodes the full distance, "not reaching" reflects how the generator renders the prompt rather than the true geometry; if u(g)u(g) encodes only the horizontal distance and ignores the height difference (the UAV starts on a rooftop), the prompt itself omits height information. In either case, the discriminative signal in this example does not necessarily reflect whether the candidate is actually reachable.

I suggest adding a direct candidate-level analysis. Visit-All already flies to every narrowed candidate and scores each observed flight with the Backtracking judge (p. 9), so without new flight experiments the authors can report the agreement between Foresight scores and actual-flight scores or outcomes (e.g., top-1 agreement or rank correlation), as well as the agreement between predicted and actual flights on key action events. This would clarify when the system benefits from video prediction and when it relies on Backtracking to recover.

4. There is a clear trade-off between endpoint success and trajectory fidelity, and direct evidence for the value of process verification is lacking

On AerialVLN-Fine, FaB-Fly improves SR from 8.0 (FineCog-Nav) to 15.1, but nDTW drops from 22.5 to 9.1 and NE increases from 91.4 to 97.6 (p. 7, Table 2). This trade-off is not limited to AerialVLN-Fine: on AerialVLN-S val-unseen, FaB-Fly's OSR (28.3) is clearly lower than that of LookasideVLN (36.0) and CityNavAgent (35.2), and its NE (79.7) is higher than CityNavAgent's (60.2).

The authors attribute the lower nDTW to the low-level controller flying directly toward landmarks rather than closely tracing the reference path, and to return flights reducing trajectory similarity, and they acknowledge this limitation in the conclusion (p. 8, lines 414–425; p. 9, lines 482–485). However, Table 3 shows that return flights explain only a small part of the gap: w/o Both, which involves no return flights, already has an nDTW of only 10.4, below FineCog-Nav (22.5) and even NavGPT (15.9), and adding Backtracking lowers it by only a further 1.0–1.4 (10.4→9.410.4 \rightarrow 9.4 and 10.5→9.110.5 \rightarrow 9.1). The main gap stems from the base framework itself.

A lower nDTW does not directly imply violating the language instruction, since the reference path may not be unique and return flights also increase trajectory differences. However, the core motivation of this paper concerns process requirements such as crossing a river, detouring, and turning, and endpoint SR and average NE alone cannot distinguish among the following cases:

  • reaching the target while satisfying the action constraints in the instruction;
  • reaching the target via a path that does not meet the instruction's requirements;
  • satisfying the instruction with a trajectory whose shape differs from the reference route. I suggest a segment-level analysis based on the existing trajectories that distinguishes semantic action violations from general path deviations. This would verify the problem the paper targets more directly than explaining the drop in nDTW.

5. The cost comparison is incomplete, and the source of the additional overhead is hard to interpret

The timing in Figure 4 already includes flight time: the full method takes about ×1.20\times 1.20 the time of w/o Both while improving SR by 0.8 points, and Visit-All takes about ×1.86\times 1.86 the time of the full method while achieving about 0.7 points higher SR (p. 9). These numbers support FaB-Fly's cost advantage over exhaustive visiting.

However, Figure 4 omits w/o Foresight. The SR of this variant is not only the highest in Table 3 but also higher than that of Visit-All (15.1+0.7=15.815.1 + 0.7 = 15.8 according to the text): a variant that generates no videos, first flies to the first location of the lowest-cost route, and retries at most once outperforms Visit-All, which physically flies to every candidate. This makes the omission of w/o Foresight all the more worth questioning, and suggests two possibilities: selecting among all candidates by Backtracking scores may not be reliable, or these differences lie within noise.

The w/o Foresight variant has worse NE, and its total time also depends on how often Backtracking is triggered, so it should be measured under the same timing protocol. If the main value of the full system lies in reducing NE, the cost discussion should present this benefit explicitly.

In addition, each segment may require generating a 5–10 s video and calling the judge for each of up to 8 candidates, yet the total time increases by only about 20% over w/o Both. Because Appendix A.3 deduplicates routes by their final endpoints rather than by the current candidates, the number of current candidates entering Foresight may be substantially smaller than 8, and the correct current candidate may also be excluded.

I suggest reporting the distribution of the number of candidates per segment, the fraction of single-candidate segments, the fraction of segments in which the correct candidate is retained, and the number of videos actually generated per episode, along with the hardware or service environment, whether calls run serially or in parallel, and how video generation, judge calls, and flight time are each timed. This would help reproduce the timing comparison and show in how many decisions Foresight actually participates in a multi-candidate choice.

6. Several key implementation details are missing, which affects reproducibility and the attribution of results

The paper specifies the models used, the prior-survey procedure, the main parameters of candidate narrowing, and the single-retry rule, but the main text and Appendix A still lack the following:

  • the value of the backtracking threshold τ\tau (a location is accepted when zobs≥τz^{\mathrm{obs}} \geq \tau) and how it was chosen;
  • the weight λ\lambda in the route cost Γ(r)\Gamma(r) (Eq. (4)), and the mapping from motion phrases other than "turn left"/"turn right" to the reference turn angle ϕ(mj)\phi(m_j) (Appendix A.3; only ϕ(turn left)=+45∘\phi(\text{turn left}) = +45^\circ and ϕ(turn right)=−45∘\phi(\text{turn right}) = -45^\circ are given);
  • the text-embedding model and similarity threshold used for retrieval (Appendix A.2), and the distance threshold for merging landmark detections (Appendix A.1);
  • how the MLLM judges' outputs are mapped to the scores z^(g)∈[0,1]\hat{z}(g) \in [0, 1] and zobs∈[0,1]z^{\mathrm{obs}} \in [0, 1] (Eqs. (2)–(3)), the complete judge prompts, and the frame-sampling settings for the predicted and observed videos;
  • whether videos are still generated when only one candidate is available;
  • the control implementation and arrival criterion for returning to the departure pose qdepq_{\mathrm{dep}}, and whether return flights count toward the action budget;
  • how w/o Foresight selects the alternative candidate after Backtracking is triggered. The last item particularly needs to be specified. Appendix A.3 states that, in the w/o Foresight variant, the UAV flies to the first location of the lowest-cost route, but the retry in the full method relies on the Foresight scores, g=arg⁡max⁡g∈C~i∖{gF}z^(g)g = \arg\max_{g \in \tilde{C}_i \setminus \lbrace g^F \rbrace} \hat{z}(g) (Algorithm 1, line 8). It is unclear which ranking replaces z^(g)\hat{z}(g) once Foresight is removed, and this variant happens to achieve the highest SR.

One minor clarification remains. The related-work section states that FineCog-Nav judges subgoal completion by combining observations, subgoal memory, and imagined scene descriptions, but the introduction characterizes it as a check based on the observation upon arrival. I suggest that the introduction distinguish more precisely between "verifying directly with the flight video" and "judging completion using historical memory," so that readers do not take the difference to be whether historical information is used at all.

Questions for authors

1. Which metrics does the benefit of combining the components target, and by what criterion was the default configuration chosen?

Please specify which metrics "complementarity" refers to and the criterion used to choose the default configuration. Relative to w/o Foresight, the full system improves average NE (111.4→97.6111.4 \rightarrow 97.6) while SR decreases (16.3→15.116.3 \rightarrow 15.1); can failure cases or the distribution of endpoint errors explain this? Please report the number of repeated runs, the uncertainty ranges, and the number of episodes in each evaluation split.

2. How much of the improvement over external baselines comes from the new modules?

Could the authors add w/o Both on AerialVLN-S, as well as a baseline on AerialVLN-Fine that also uses environmental priors? Please state the source of each baseline result in Tables 1 and 2 (cited directly or re-run), the actual model configurations, and the environmental priors used. Please also specify in the implementation details how the distance for SR is computed (2D or 3D), the success threshold, and the evaluation implementation.

3. Can video prediction reliably support ranking candidates for real flights?

Could the authors use the existing Visit-All records to report the agreement between Foresight scores and actual-flight scores or outcomes, especially when the target is not currently visible, is far away, or the instruction requires detouring or crossing? In addition, does u(g)u(g) encode the height difference? Please use this to explain why, in Figure 3(a), the predicted flight for the 10 m prompt does not reach the second candidate.

4. Which specific instruction-following behaviors does process verification improve?

Could the authors use the existing trajectories or a segment-level analysis to distinguish reasonable path deviations, the effect of return flights, and genuine violations of action constraints, so as to show more directly how Foresight and Backtracking affect instruction following?

5. Where does the additional overhead come from, and what does w/o Foresight cost?

Could the authors report the time of w/o Foresight under the same timing protocol as Figure 4, together with the distribution of the number of candidates per segment, the fraction of single-candidate segments, the fraction of segments in which the correct candidate is retained, the number of videos actually generated per episode, and the parallelization scheme? If the full system is chosen mainly to reduce NE, I suggest making this trade-off explicit in the cost discussion.

6. Could the authors provide the core implementation details and ablation definitions?

Please provide the items listed in Weakness 6, especially τ\tau, λ\lambda, the retry rule of w/o Foresight, and the implementation of returning to the departure pose. Please also clarify which aspects of the action space, termination conditions, and action budget follow the existing benchmarks and which were modified.

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

FaB-Fly is a hierarchical zero-shot framework for aerial VLN that scores candidate flights against the current instruction segment, both before and after execution. Candidates retrieved from the landmark memory are first filtered by geometric rules. Foresight generates, for each candidate, a flight video via a video world model, conditioned not on the evaluated segment but on the candidate's own flight instruction, and an MLLM-based judge scores the prediction. Backtracking assesses a posteriori the observed flight after execution and resumes with the next-best candidate if the score is low. On AerialVLN-S the reported SR is 18.6/15.3 (LookasideVLN: 14.7/12.6); on AerialVLN-Fine it is 15.1 (FineCog-Nav: 8.0).

Critical strengths

The design that scores candidate flights against the instruction segment, both before and after the flight, usefully decomposes errors in hierarchical navigation; conditioning the video model on u(g) rather than on the evaluated segment, described in Section 3.3, reduces the risk that the generator merely replays a motion prescribed for each candidate. The baselines are relevant comparison partners: LookasideVLN shares the landmark-memory assumption, and the paper places CityNavAgent and LookasideVLN in the same group as FaB-Fly, as methods that exploit information about the surroundings in advance. FineCog-Nav is relevant through its verification of subgoal execution. The limitations section honestly acknowledges the drop in nDTW on AerialVLN-Fine (9.1, versus 22.5 for FineCog-Nav, Table 2).

Critical weaknesses

1.The evidence for the components is thin, and some of it contradicts the thesis. In Table 3, removing Foresight yields the best SR (16.3 versus 15.1 for the full model), and the full model exceeds the skeleton without either of the two modules by only +0.8 SR (14.3→15.1). In Table 4, giving the video to a single judge lowers SR (11.7/13.0) below the purely image-based variant (14.4). Neither uncertainty over repeated runs nor decoding settings are reported, so it is difficult to determine whether the +3.9/+2.7 SR over LookasideVLN stems from the proposed mechanisms, from run-to-run variance, or from other differences between frameworks and memories. The trade-off with trajectory following is also heavy: FineCog-Nav achieves an nDTW of 22.5 and an NE of 91.4, while FaB-Fly achieves 9.1 and 97.6, even though following prescribed trajectories is precisely the paper's motivation.

2.The frequency with which the generated flights mislead the judge, and the frequency with which they help it, are not measured, even though the discrepancy between predicted flight and executed flight is precisely the failure mode that motivates Backtracking. Moreover, the landmark memories for val-unseen and AerialVLN-Fine are constructed from images "captured across each scene" (a prior sweep of the environment), and the paper does not state whether the coverage behind the reported LookasideVLN/CityNavAgent values on val-unseen matches this imagery. The total time per episode is reported in Fig. 4, but the cost of each decision — generating up to 8 candidate videos, scoring each with the Foresight judge, Backtracking calls — is not broken down, and there is no cost-benefit comparison with the simple variant without Foresight, which yields a higher SR.

Questions for authors

1.Please report the decoding settings, and, where sampling is stochastic, repeat Tables 3 and 4 under controlled generation and fixed judge seeds, with means and standard deviations. Why does adding Foresight to the variant with Backtracking lower SR from 16.3 to 15.1, while NE improves? Does this tendency persist across repetitions?

2.Can the precision of the predictions be measured, for example by having the judge assess the observed flight against the chosen candidate's predicted flight? And what proportion of Backtracking rejections correspond to misleading prediction videos?

3.Does the image coverage behind the reported values for LookasideVLN and CityNavAgent on val-unseen correspond to the "captured across each scene" construction that FaB-Fly uses?

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

Related papers

Discussion (0)

Sign in to comment.