FlowPilot: Real-Time World-Action Modeling for Agile UAV Navigation
1. Paper overview
In one sentence: FlowPilot jointly denoises future-depth latents and a compact polynomial trajectory to improve agile navigation, while its onboard timing and feasibility claims require narrower interpretation than its headline suggests. e-probleme-architecturee-traininge-simulatione-ablatione-depthe-latencye-flight
| At a glance | What to know |
|---|---|
| Research problem | Source description Fast flight through unknown clutter requires both low perception-to-action delay and a reference with useful velocity, acceleration, and jerk. The authors target the gap between map-dependent optimization and reactive learned policies, using anticipated depth geometry to guide a compact trajectory generator. e-problem |
| Core mechanism | Source description A dual-stream mixture-of-transformers (MoT) jointly denoises future-depth and action latents, exchanging information through shared attention while preserving modality-specific parameters. e-architecture |
| A key reported result | Synchronous future-depth denoising in PX4 SITL: Full: 4.0%; 12.98 ± 1.29 s; 4.14 ± 0.32 m/s. Collision rate; task time; mean speed. Full model versus future-depth latents held fixed as noise while action latents denoise; table-specific trial count and commanded speed are unstated. Depth-frozen: 26.0%; 14.17 ± 1.51 s; 3.83 ± 0.34 m/s. The collision reduction is 22 percentage points. The meaning of ± is unspecified. Normalized jerk is 65.4 ± 31.9 versus 63.9 ± 41.2, so the table does not establish improvement in every metric. e-ablation |
| Reading caution | Source description The authors identify short-horizon, single-depth-camera navigation as vulnerable to thin obstacles, transparent surfaces, and unreliable sensing range. e-limitations |
Core contributions
- Source description
A dual-stream mixture-of-transformers (MoT) jointly denoises future-depth and action latents, exchanging information through shared attention while preserving modality-specific parameters. e-architecture
- Source description
A state-constrained degree-7 Bernstein curve compresses the generated action to five three-dimensional control points and provides analytic derivatives. e-bernstein
- Source description
The authors demonstrate closed-loop simulation and onboard indoor/forest flights; these are executed navigation results, distinct from the qualitative depth predictions. e-simulatione-flighte-depth
Figure 2. Two expert streams exchange information inside each denoising layer. Original paper, p. 3 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read upward from the inputs. The blue side supplies current-depth conditioning and future-depth information; the orange side supplies state and goal, speed, previous trajectory, noisy free control points, and register tokens. Each side has its own projection and feed-forward network, but the horizontal attention block joins their tokens. The separate flow-time inputs modulate the two streams. Snowflake symbols on the Wan encoder and decoder agree with the frozen VAE described in Section III-B; the MoT and action pathway are learned. The five labeled output points are the free part of the polynomial, not the entire reference or motor commands. e-architecturee-inputse-bernsteine-flowe-ablation
What it supports. Future-scene and action representations can influence one another within a denoising step. This supports joint world-action prediction without requiring identical parameters for both modalities. At deployment, the controller needs the resulting curve; decoded future images are optional diagnostics, while latent-depth updates have a separate tested role.
Where the evidence stops. The artwork places noisy-depth thumbnails below the encoder, whereas Equations (8)–(11) define noise interpolation on VAE latents. Use the equations for corruption placement. The diagram also does not specify a complete deployed token/noise schedule.
2. Motivation
2.1 The problem and the proposed response
Fast flight through unknown clutter requires both low perception-to-action delay and a reference with useful velocity, acceleration, and jerk. The authors target the gap between map-dependent optimization and reactive learned policies, using anticipated depth geometry to guide a compact trajectory generator. e-problem
2.2 What this reading follows
A drone needs more than a plausible next waypoint: its controller needs a smooth reference quickly enough to react to nearby clutter. FlowPilot addresses both requirements by coupling future-depth prediction with a small generative action space. Read the architecture and training masks first to see where geometry and motion exchange information. Then separate three kinds of evidence: simulated goal-reaching success, a targeted latent-depth intervention, and onboard timing. The depth rollout explains what the model predicts, but the collision table tests a control consequence. Physical indoor and forest flights establish deployment examples, while missing protocol details and timing boundaries limit broader conclusions. e-probleme-architecturee-traininge-simulatione-ablatione-depthe-latencye-flight
3. Research context
We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.
| Catalog dimension | Recorded classification |
|---|---|
| Major category | WAMs |
| Architecture | Dual-system |
| Prediction paradigm | Joint prediction |
| Quadrant | Q3 · Dual-system × Joint prediction |
3.1 Evidence-based assessment
Supports the recorded classification
Joint prediction is directly supported by coupled future-depth and action velocity fields. Distinct video/action expert weights support the recorded Dual-system label at the expert level, although both live inside one integrated MoT rather than a serial world-model/planner pipeline. Neither joint training nor the authors' 'single model' wording alone establishes a shared-weight One Model architecture. e-architecturee-flow
This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.
4. Problem formulation
4.1 Inputs and outputs
| Inputs | Outputs |
|---|---|
|
|
4.2 Equations and their role
5. Method in detail
5.1 Trace a replan from depth to an executed reference
The policy starts with depth and a state–goal description in the drone's current body frame. During joint modeling, video and action experts each predict a velocity field, and shared attention lets the partially denoised representations condition one another. This is the central causal path to examine: geometry is not simply decoded and handed to a separate optimizer. A flow sample supplies five free control points; the remaining points come from current state, and the resulting curve goes to OMMPC. Deployment can skip future-image decoding while retaining informative latent-depth computation. Table I tests the latter by freezing future latents as noise, which differs from removing the video expert altogether. The observed collision difference supports this coupling under the reported protocol, without proving that every predicted depth frame is metrically accurate. e-inputse-architecturee-flowe-bernsteine-setupe-ablatione-depth
Figure 4. Broad depth pretraining precedes joint alignment with expert trajectories. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the left pyramid: quantity increases toward the 16 h simulation base, while quality points toward the 2 h real-data apex. These are the authors' organizing axes, not measured quality scores. On the right, rows are queries and columns are keys. In phase 1, the filled three-by-three block joins conditioning-frame tokens V_c, future-video tokens V, and state–goal token P. Hatching masks speed S, previous trajectory T_p, and action A in both directions. Phase 2 fills the whole matrix. The masks and their direction agree with the caption and Section III-F: action information cannot enter phase-1 world modeling through the masked tokens. e-datae-traininge-repro
What it supports. The curriculum separates a broad navigation representation from expert action alignment. The second phase can exchange geometry and motion information directly because all token groups attend jointly. Its SUPER action targets come specifically from forested Flightmare environments, so the depth pyramid does not imply equally broad expert-action coverage.
Where the evidence stops. The paper provides no ablation that separately establishes the value of each data tier or the two-stage curriculum. It also omits phase-specific sample counts, optimization schedules, and a complete train/test split.
5.2 Separate what the two training phases can learn
Phase 1 learns scene evolution from the depth pyramid, with attention confined to current-frame, future-video, and state–goal tokens. The absent action pathway in the mask matters: expert action alignment is introduced later, rather than inferred from the presence of real depth in the pretraining set. Phase 2 enables all-token attention and supervises both velocity fields, using SUPER trajectories collected in forested Flightmare. Independent noise levels expose the network to differently corrupted modalities, including clean-action/noisy-video and clean-video/noisy-action regimes. This supplies a rationale for varying depth corruption at inference, but does not itself certify an unchanged closed-loop distribution. Reproducing the curriculum also requires details absent from the paper, particularly loss weighting, action normalization, phase schedules, and precise partitions of the available data. e-datae-traininge-flowe-ablatione-repro
5.3 Separate curve smoothness from replan consistency and feasibility
A degree-7 Bernstein curve has eight control points. Fixing the first three to satisfy initial position, velocity, and acceleration leaves fifteen scalar outputs, arranged as five three-dimensional points. This construction provides analytic velocity, acceleration, and jerk, avoiding numerical differentiation of independent waypoint predictions. It does not impose a bound on those derivatives or guarantee clearance from obstacles. A second issue arises across replans: independent samples can switch maneuvers even when each curve is smooth. The previous-trajectory token softly biases the next sample, with dropout teaching the model to operate without that token. As reader analysis, the two devices address different problems: polynomial structure supports within-reference smoothness, whereas conditioning encourages temporal consistency. Neither alone proves exact continuity during tracking errors or dynamic feasibility at all vehicle states. e-bernsteine-previouse-limitations
5.4 Training and inference
During training
World-model pretraining uses the depth pyramid: 16 h labeled IsaacLab, 8 h Flightmare, and 2 h onboard data. Only conditioning-frame, future-video, and state–goal tokens attend; speed, previous-trajectory, and action tokens are masked. e-datae-training
World-action alignment enables all-token attention and both velocity losses. Action supervision comes from the LiDAR-based SUPER planner, collected only in forested Flightmare environments. This differs from the unspecified coarse-coverage optimizer used for phase 1. e-training
Independent video/action noise levels train multiple conditioning regimes. Equation (9) shifts noise toward the high-noise regime with shift factor 5; the loss notation still writes uniform noise sampling, without explicitly showing where the shift is substituted. e-flow
The previous trajectory's eight control points form a soft conditioning token. Training replaces it with a learned null embedding with probability 0.2. The authors report copying or horizon collapse when warm-starting, but give no quantitative ablation of that choice. e-previous
During inference
The default decoder uses three Euler denoising steps over a 1.6 s trajectory horizon. Each replan starts from noise. Deployment consumes action control points without decoding future-depth images; the depth-frozen experiment nevertheless tests a benefit from updating future-depth latents during action generation. e-setupe-previouse-architecturee-ablation
Onboard depth is resized to 160 × 96 from a RealSense D455, whose stated maximum range is 6 m. VINS-Fusion estimates state, TensorRT executes the neural model, host code reconstructs the curve, and OMMPC tracks the latest reference at 100 Hz. That controller frequency is not the model's demonstrated inference rate. e-setupe-latency
5.5 Implementation flow
- Express the task egocentrically
The frame is anchored at current position and heading. The state–goal vector contains current position, velocity, acceleration, jerk, orientation quaternion, and goal position, velocity, and acceleration. e-inputs
- Couple two denoising streams
A frozen Wan2.2 VAE encodes one conditioning depth frame and eight future frames; video tokens use RoPE-3D. Action tokens encode the task conditions, noisy free control points, previous trajectory, and register tokens. Separate projections and feed-forward networks surround joint attention, modulated by independent video/action flow times. e-architecture
- Construct a reference
Recover the first three control points from measured position, velocity, and acceleration; predict the other five in normalized space. Evaluate the polynomial and its analytic derivatives. Smoothness and initial-state consistency do not enforce collision avoidance or vehicle limits. e-bernstein
6. Experiments & results
FlowPilot couples future-depth and trajectory denoising for onboard quadrotor navigation. Its action is five free Bernstein control points, converted into a state-consistent reference for a separate tracking controller. Simulated comparisons and physical flights support the approach, while the strongest mechanism test freezes future-depth latents during inference. Image decoding is optional; this does not establish that latent depth computation can be removed. The reported latency requires careful accounting.
6.1 Read the original evidence
Figure 6. Success stays high for FlowPilot as commanded speed increases in these simulated forests. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each panel at a fixed commanded velocity before comparing methods. Panel (a) is sparsity 10, the denser forest; panel (b) is sparsity 15. The horizontal axis is commanded speed, and the vertical axis is goal-reaching success across 20 trials per speed and sparsity. Red circles denote FlowPilot, magenta triangles its action-only variant, blue squares EGO-Planner, and yellow crosses YOPO. Keep the two panels separate because obstacle density changes the task. Also distinguish this action-only variant, which removes the future-depth stream and shared attention, from Table I's intervention that freezes future-depth latents during inference. e-simulatione-setupe-ablation
What it supports. At 8 m/s, the plotted FlowPilot success rate is 90% in both panels, compared with YOPO's 10% in dense and 35% in sparse forests and EGO-Planner's 0% in both. The separation grows under challenging conditions; lower-speed points include ties between FlowPilot and EGO-Planner.
Where the evidence stops. The curves provide no confidence intervals or per-trial outcomes. Twenty trials per point support a finite-sample comparison, not a universal safety rate. Commanded speed does not establish sustained flight at that speed.
Table II. The labeled inference path is fast, but the complete timing boundary remains unresolved. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the columns as mean, standard deviation, and maximum milliseconds; unlike Table I, this table defines the uncertainty column. Three denoising steps average 14.924 ms, and the distilled encoder averages 1.163 ms. The row labeled total inference reports 16.294 ms mean and 17.055 ms maximum. Preprocessing and MPC appear as separate rows. Do not automatically add all maxima or assume every row belongs to a serial total. Section IV-C specifies TensorRT and host-side Bernstein reconstruction, while Section IV-A separately states 100 Hz trajectory tracking. That tracking rate refers to the downstream controller, not a measured model update frequency. e-latencye-setup
What it supports. The labeled neural inference total is below 18 ms on the reported Jetson. As reader arithmetic, preprocessing plus encoding plus denoising means sum to 20.834 ms, already above that total. The difference calls for explicit timing boundaries or overlap information before accepting the abstract's full perception-to-action budget.
Where the evidence stops. The source does not explain asynchronous overlap, timestamp boundaries, or whether preprocessing and reference reconstruction are included. The arithmetic sum is a diagnostic consistency check, not a measured end-to-end latency; MPC is reported separately.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Closed-loop forest navigation at 8 m/s commanded speed Workstation simulation; sparsity 10 (dense) and 15 (sparse); 20 trials per velocity and sparsity; matched scenes, start–goal pairs and criteria. | Figure 6 plots 90% for FlowPilot in both scenarios at 8 m/s. Goal-reaching success rate | At 8 m/s, YOPO plots 10% dense and 35% sparse; EGO-Planner plots 0% in both. These are graph-read results without confidence intervals. Commanded speed is not necessarily achieved speed; the setup does not provide a reproducible scene split. e-simulatione-setup |
| Synchronous future-depth denoising in PX4 SITL Full model versus future-depth latents held fixed as noise while action latents denoise; table-specific trial count and commanded speed are unstated. | Full: 4.0%; 12.98 ± 1.29 s; 4.14 ± 0.32 m/s. Collision rate; task time; mean speed | Depth-frozen: 26.0%; 14.17 ± 1.51 s; 3.83 ± 0.34 m/s. The collision reduction is 22 percentage points. The meaning of ± is unspecified. Normalized jerk is 65.4 ± 31.9 versus 63.9 ± 41.2, so the table does not establish improvement in every metric. e-ablation |
| Onboard neural inference latency Jetson Orin NX Super 16 GB, TensorRT, three denoising steps. | 16.294 / 0.180 / 17.055 ms. Reported total inference mean / standard deviation / maximum | Denoising alone averages 14.924 ms; distilled encoding 1.163 ms; preprocessing 4.747 ms; MPC 2.384 ms. The total's inclusion boundaries and overlap are unexplained. It cannot safely substantiate a serial, complete sensing-to-control budget below 18 ms. e-latencye-setup |
| Physical indoor and forest navigation Fully onboard flights with depth/state feedback, no prebuilt navigation map; indoor layouts and hilly natural forest. | Indoor peaks: 3.8 and 3.1 m/s. A 100 m forest traversal reaches 5.5 m/s in sparse regions and 4 m/s in dense regions. Reported attained speeds and route length | No matched physical baseline or aggregate flight-success distribution is reported. These examples support physical feasibility under the tested conditions, not universal transfer to arbitrary obstacles. e-flight |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table I. Updating future-depth latents improves collision and traversal outcomes in the reported ablation. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Compare columns within a row. Full performs synchronous future-depth and action denoising; Depth-frozen holds future-depth latents fixed as noise while continuing to denoise the action. The collision row therefore probes whether evolving latent predictions contribute during action generation. Next read task time and speed together: lower traversal time accompanies higher mean speed in Full. The last row matters because it prevents treating every number as a uniform win: normalized jerk is slightly higher for Full. Retain the printed plus/minus quantities literally; the paper does not define them as standard deviations, standard errors, or confidence intervals. e-ablatione-flow
What it supports. Full reports 4.0% collisions versus 26.0%, a reduction of 22 percentage points. Task time falls from 14.17 ± 1.51 s to 12.98 ± 1.29 s while mean speed rises from 3.83 ± 0.34 to 4.14 ± 0.32 m/s. These results support useful anticipatory information in latent updates under this test.
Where the evidence stops. This table's trial count, exact commanded speed, uncertainty convention, speed-ratio definition, and jerk normalization are unspecified. Independent training noise levels motivate the intervention but do not prove identical closed-loop state distributions or a theoretical imitation-performance ceiling.
Figure 8. A qualitative rollout connects future geometry to the trajectory horizon. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Compare the two rows vertically at the same time label. The upper row is predicted depth and the lower row is ground truth; the shared current frame anchors the sequence. Later predictions retain broad bright and dark regions but are visibly noisier and less sharply structured than the corresponding observations. The lower panel locates the conditioning frame along a red trajectory and labels the 1.6 s horizon. It also marks a 0.2 s interval near the start; the displayed frame pairs above use 0.4 s spacing. These annotations show temporal alignment without supplying a quantitative depth-accuracy measure. e-depthe-ablatione-architecture
What it supports. The figure illustrates the kind of future geometry available to the action representation and gives qualitative context for Table I. It is consistent with the authors' claim that obstacle layout and ego-motion trends can inform action denoising, even when deployment does not decode the frames for the controller.
Where the evidence stops. One selected rollout cannot establish prediction calibration, obstacle-clearance accuracy, or causal control benefit by itself. The noisy images should not be read as a certified free-space map; the closed-loop ablation supplies different evidence.
7. Analysis & limitations
7.1 What the evidence leaves open
The authors identify short-horizon, single-depth-camera navigation as vulnerable to thin obstacles, transparent surfaces, and unreliable sensing range. e-limitations
The conclusion calls trajectories dynamically feasible by construction, whereas Section III-C explicitly makes feasibility depend on controller and vehicle limits. The narrower method statement is warranted. Initial-state matching also does not prove perfect continuity between imperfectly tracked replans. e-bernsteine-limitations
Depth-frozen inference is not the separately trained action-only baseline. Its degradation supports useful latent coupling, but independent training noise levels do not prove identical rollout distributions or the authors' asserted imitation ceiling at SUPER performance. e-simulatione-flowe-ablation
7.2 Questions for discussion
- Does latent-depth updating retain its safety benefit after matching wall-clock planning delay?
- How often do smooth predicted curves violate vehicle limits or produce reference jumps during imperfect tracking?
8. Reproducibility audit
8.1 Requirements and known gaps
A reproduction needs the depth pyramid, SUPER forest trajectories, body-frame transforms, normalized control-point targets, two training masks, and the onboard controller stack. Missing essentials include model size, optimizer/schedule, loss weight, action normalization, data splits, and the depth-encoder distillation recipe. e-datae-inputse-traininge-reproe-setup
Proposed checks should hold scenes and action noise fixed when changing latent-depth updates, and timestamp preprocessing through reference delivery separately from MPC. Report collision uncertainty and exact timing inclusion boundaries. e-ablatione-latency
8.2 Proposed reproduction checks
The following checks are proposals motivated by the paper. They have not been run as part of this reading.
Check 1: Test latent-depth guidance with matched noise and planning delay
Reader-proposed experiment, not performed: use one joint checkpoint in PX4 SITL and paired forest scenes, start–goal pairs, speeds, and action-noise seeds. Compare synchronous depth/action updates with the paper's depth-frozen intervention, retaining the same polynomial representation and controller. First use identical denoising-step counts; then equalize reference-delivery delay so any compute difference cannot change reaction time. Predeclare trial counts, collision criteria, and uncertainty estimates, and log traversal time and normalized jerk with an explicit definition. If the collision advantage persists under paired seeds and matched delay, that strengthens the case for informative latent geometry. If it disappears, the original attribution needs qualification. Include the separately trained action-only model as an additional comparison, without treating it as the same intervention. e-ablatione-simulatione-setupe-flow
Check 2: Reconcile the Jetson timing budget with explicit end points
Reader-proposed experiment, not performed: on the specified Orin NX platform, fix input resolution, three Euler steps, and the 1.6 s horizon. Timestamp depth arrival, preprocessing completion, encoder completion, denoising completion, reconstructed-reference availability, and controller consumption. Collect repeated warm-run measurements with GPU synchronization only at clearly documented measurement boundaries, and compare the normal overlapping pipeline with a forced serial diagnostic run. Keep optional future-image decoding disabled in the primary comparison. Report per-stage distributions and observed end-to-end latency separately from the 100 Hz controller period. If preprocessing is included and the actual arrival-to-reference path remains below 18 ms, the broad budget is supported; if only the neural subpath does, Table II's narrower interpretation is confirmed. e-latencye-setupe-architecturee-bernstein
8.3 Reading coverage
Visual audit: All eight supplied PDF pages were rendered and visually inspected: p. 1 title/authors/version/affiliations; p. 2 overview; pp. 3–5 architecture, all method equations, smoothness schematic, data pyramid and attention masks; p. 6 conditioning, platform, simulation protocol and Figures 5–7; p. 7 Tables I–II, future-depth rollout, forest flights and deployment details; p. 8 flight continuation, conclusions and references. All six final crops were individually viewed at their original 300-DPI render resolution. Architecture arrows were checked against the caption and latent-noise equations; the schematic-versus-equation corruption placement is disclosed. Figure 4 query/key directions and filled/masked token blocks agree with its caption. All numerical, method and proposed-check evidence pages are included here. No appendix is present, and no separate supplement or external code was inspected.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8. Appendix coverage: not present.
Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.
Text reading scope & known omissions
- PDF pp. 1–2: Abstract and I. Introduction
- PDF pp. 2–3: II. Related Work, A–D
- PDF pp. 3–6: III. Method, A–F, including all equations and training phases
- PDF pp. 6–8: IV. Experiments, A–C
- PDF p. 8: V. Conclusions and References
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- Identity: the title and all nine authors match the catalog. The title page identifies arXiv:2608.00635v1 [cs.RO], 1 August 2026. No other revision was supplied or compared; this report covers that version only.
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout. This acquisition limitation was addressed by visually inspecting all eight supplied PDF pages and every final crop.
- Separate supplemental material availability has not been fully verified.
- No separate supplement was supplied. Code and external resources were not inspected; no experiments were reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e-identityPDF p. 1, title, author block, affiliation footnote and arXiv margin
Exact title; nine catalog-matching authors; arXiv:2608.00635v1 [cs.RO], 1 Aug 2026. Affiliations are Beijing Institute of Technology, Zhongguancun Academy, Shanghai Jiao Tong University, and The University of Hong Kong.
Go to primary source ↓e-problemPDF pp. 1–2, I. Introduction and Figure 1 caption
The problem combines onboard timing, future-scene prediction and controller-trackable trajectories; the deployment controller consumes trajectories.
Go to primary source ↓e-inputsPDF p. 3, III-A, Equations (1)–(2)
Body-frame depth, 25-dimensional state–goal vector, speed and previous trajectory condition future-depth and 5 × 3 action latents.
Go to primary source ↓e-architecturePDF p. 3, Figure 2 and III-B(a–c); p. 4, III-B(c–d)
Frozen Wan2.2 VAE, one conditioning and eight future frames, RoPE-3D, modality-specific projections/FFNs, joint attention and independent flow times; future-image decoding is unnecessary in deployment.
Go to primary source ↓e-bernsteinPDF p. 4, III-C, Equations (3)–(7) and Figure 3
Degree-7 Bernstein position curve; first three points fixed from state; five predicted points; analytic derivatives. The method explicitly leaves dynamic feasibility dependent on tracking and vehicle limits; Figure 3 is schematic.
Go to primary source ↓e-dataPDF p. 4, III-D; p. 5, Figure 4(a)
Depth pyramid labels 16 h IsaacLab, 8 h Flightmare and 2 h real-world depth; quantity grows downward and quality upward.
Go to primary source ↓e-flowPDF p. 5, III-E, Equations (8)–(12)
Linear noise interpolation, target epsilon minus clean latent, shifted schedule with factor 5, independent modality noise levels and weighted squared velocity losses; lambda has no stated value.
Go to primary source ↓e-trainingPDF p. 5, Figure 4(b), caption and III-F(a–b)
Phase 1 permits attention among V_c, V and P only. Phase 2 enables full attention and both losses, using SUPER action targets from forested Flightmare.
Go to primary source ↓e-previousPDF pp. 5–6, III-F(c)
Previous eight control points are transformed to the current body frame and used as soft conditioning, dropped to a null token with probability 0.2. Authors describe warm-start copying/collapse without numerical comparisons.
Go to primary source ↓e-setupPDF p. 6, IV-A and Figure 5(a)
Workstation: i9-14900K/RTX 4090. Default three Euler steps and 1.6 s horizon. Onboard Orin NX Super 16 GB, D455 depth at 160 × 96, 6 m stated range, VINS-Fusion and 100 Hz OMMPC.
Go to primary source ↓e-simulationPDF p. 6, IV-B and Figure 6(a–b), 8 m/s points and caption
20 trials per speed/sparsity; matched scenes and criteria. FlowPilot: 90% both panels; YOPO: 10% dense, 35% sparse; EGO-Planner: 0% both. Action-only removes the video stream and shared attention but retains Bernstein output.
Go to primary source ↓e-ablationPDF p. 7, Table I, Full/Depth-frozen columns, and IV-B paragraphs below Figure 8
PX4 SITL depth-frozen intervention holds future latents as noise. Collision 4.0/26.0%; time 12.98 ± 1.29/14.17 ± 1.51 s; mean speed 4.14 ± 0.32/3.83 ± 0.34 m/s; normalized jerk 65.4 ± 31.9/63.9 ± 41.2. Trial count, ± convention and jerk normalization are unspecified.
Go to primary source ↓e-depthPDF p. 7, Figure 8 and final IV-B paragraph
Predicted versus ground-truth depth at current time and 0.4 s intervals through 1.6 s; qualitative obstacle-layout and ego-motion evidence, with visibly noisy predictions and no quantitative depth-error metric.
Go to primary source ↓e-latencyPDF p. 7, Table II, all rows, and IV-C opening paragraph; p. 1, Abstract
Table lists preprocessing 4.747 ms, encoder 1.163 ms, three-step flow 14.924 ms, total 16.294 ± 0.180 ms (maximum 17.055), and MPC 2.384 ms mean. TensorRT and host reconstruction are stated; component overlap and total inclusion are not explained despite the abstract's full-pipeline sub-18 ms claim.
Go to primary source ↓e-flightPDF p. 6, Figure 5(b–c); p. 7, Figure 9 caption and IV-C; p. 8, IV-C continuation
Two indoor layouts reach 3.8/3.1 m/s. Figure 9 describes 80 m and 100 m forest flights; the latter reaches 4 m/s dense and 5.5 m/s sparse. Colored maps are only visualization, unavailable to navigation.
Go to primary source ↓e-limitationsPDF p. 8, V. Conclusions, both paragraphs
Conclusion asserts construction-level dynamic feasibility, then identifies short horizon and single-depth-camera vulnerabilities to thin/transparent objects and sensing range; References follow with no appendix.
Go to primary source ↓e-reproPDF p. 3, Figure 2 layer count N and III-B; p. 5, III-E/F; p. 7, IV-C encoder description
No numeric layer count/model size, complete optimization settings, action normalization, train/test split or distillation procedure is provided in these method and deployment descriptions.
Go to primary source ↓8.5 Primary sources
FlowPilot: Real-Time World-Action Modeling for Agile UAV Navigation ↗
PDF · 6,468 extracted words
Source fingerprint
024ca00d7e69c53c75d32bfcd74271938f16f75f7ccb8b49fa9989c4e00c8176