Pith. sign in

REVIEW 3 major objections 5 minor 32 references

AeroAct claims that feeding nine temporally spaced reference frames to a video-diffusion transformer lifts language-conditioned quadrotor tracking and search from 20% success with 90% collisions to 100% success with 0% collisions, and that

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 00:28 UTC pith:TLJ27OPI

load-bearing objection A competent systems paper with a striking temporal-context result, but the WAM-defining visual loss is never ablated, so the central claim is plausible yet unproven. the 3 major comments →

arxiv 2607.14997 v1 pith:TLJ27OPI submitted 2026-07-16 cs.RO

AeroAct: Action-Centered World-Action Models for Language-Conditioned Quadrotor Flight

Xinhong Zhang , Qiyuan Zhu , Yubo Huang , Haolin Chen , Runqing Wang , Yuhao Mo , Zhongxin Chen , Yu Hu
show 3 more authors
Xinjiang Wang Jian Sun Gang Wang
This is my paper
classification cs.RO
keywords world-action modelquadrotor navigationvision-language-actionvideo diffusion transformerfuture-frame supervisiontrajectory chunk generationegocentric visual historylanguage-conditioned flight
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that a world-action model can be brought out of tabletop manipulation and flown on a real quadrotor. It claims that future-frame video prediction, used only as training-time supervision, can shape a policy that decodes smooth, dynamically executable trajectory chunks at inference, without generating video. The evidence is that in Isaac Lab the number of reference frames is decisive: one frame gives 20% tracking success and 90% collisions, while nine frames give 100% and zero. If true, this would make generative video models a practical action interface for aerial robots, conditioning language, vision, and proprioception into commands a low-level controller can track.

Core claim

AeroAct's central claim is that a video-diffusion transformer can be repurposed as an action-centered world-action model for flight: given a temporal stack of egocentric frames, proprioception, and a language instruction, it predicts a chunk of local fifth-order trajectory actions, while future first-person frames are used only as dense training-time supervision and are masked out at deployment. The paper argues this is the first such world-action model flown on a physical quadrotor, and that the design decision that matters most is temporal visual context: with one reference frame closed-loop tracking succeeds only 20% of the time and collides 90% of the time; with nine reference frames bot

What carries the argument

The central object is the action-centered world-action Transformer: a single 1.3B-parameter video diffusion Transformer with a blockwise causal mask that concatenates state, reference-frame, action-token, and future-frame tokens into one sequence, with language tokens as cross-attention context. The causal mask enforces that action tokens cannot see future-frame tokens while future-frame tokens can attend to actions, so the future-frame reconstruction loss L_vis can supervise the action stream during training without leaking future information at inference. The action space itself is a chunk of 24 local fifth-order trajectory segments parameterized by endpoint position, yaw, velocity and acc

Load-bearing premise

The load-bearing premise is that the auxiliary loss for predicting future frames actually shapes the action distribution in a way that helps closed-loop flight; the paper never retrains without that loss, so if it is inert the 'world-action model' claim reduces to an imitation policy on a video backbone.

What would settle it

Retrain AeroAct on the same simulation data with λ_vis = 0 (drop the visual term in Eq. 4–5) and run the closed-loop Isaac Lab tracking and search evaluation; if success stays near 100% with 0% collisions, the visual-consequence supervision is doing no work and the world-action-model contribution is falsified.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If AeroAct's central claim holds, a world-action model can be flown on a physical quadrotor, so future-frame video prediction does not have to be generated at inference for closed-loop control.
  • Temporal visual context is the key design variable: nine reference frames (2.4 s) lift tracking success from 20% to 100% and search success from 75–90% to 100%, eliminating collisions in the tested simulation episodes.
  • Training-time future-frame supervision plus a chunked fifth-order trajectory action space produces command references smooth enough for an on-manifold MPC to track in the real world.
  • The hybrid DiffAero–Isaac–3DGS data pipeline, together with the handheld odometry device, can scale flight-like egocentric interaction data without expensive autonomous flight collection; real-world finetuning preserves simulation success with zero collisions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper: the authors never ablate the auxiliary future-frame loss L_vis; if a model trained with λ_vis = 0 matched these results, AeroAct would be a video-backbone imitation policy rather than a world-action model, and the paper's central contribution would need reframing.
  • Beyond the paper: the stark one-vs-nine reference-frame gap suggests the model is mostly learning ego-motion and relative target motion from temporal context; interpolating h (e.g., 3, 5, 7 frames) and measuring closed-loop success would reveal where the benefit saturates.
  • Beyond the paper: the authors report a single real-world demonstration of short indoor flight with offboard inference; generalizing the claim to 'real-world WAM flight' would require multi-stage commands, longer horizons, and onboard execution.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper presents AeroAct, an action-centered world-action model (WAM) for language-conditioned quadrotor flight. AeroAct adapts a pretrained video diffusion Transformer to predict a local fifth-order trajectory chunk from a history of egocentric reference frames, proprioception, and language. Future first-person frames are used only as auxiliary training-time supervision through a visual-consequence loss L_vis; at deployment the model decodes actions directly, without generating video. The authors build a hybrid simulation data pipeline based on DiffAero, Isaac Lab, and 3D Gaussian splatting, plus a handheld device for real-world data collection. They also introduce inference-time self-guidance to smooth overlapping action chunks. Closed-loop Isaac Lab experiments compare one vs. nine reference frames on target tracking and object search, and a real-world flight is shown for the instruction 'fly to the yellow foam mat.' The central claims are that temporal visual context substantially improves closed-loop performance, and that AeroAct is the first WAM executed on a physical quadrotor.

Significance. If the WAM claim is fully supported, this would be a noteworthy step: it would extend world-action models from tabletop manipulation to fast, ego-centric aerial control, and the design of training-time-only future-frame prediction with a causal mask is a sensible way to reduce inference cost and avoid error accumulation. The paper also contributes a substantial data-generation pipeline, a low-cost handheld collection device, and a closed-loop real-world flight demonstration. The authors are explicit about several limitations, including short indoor trajectories and offboard inference. The main caveat is that the component most responsible for distinguishing AeroAct from a video-backbone action-diffusion policy—the visual-consequence loss—is never ablated, leaving the WAM designation conceptually under-supported.

major comments (3)
  1. [§3.2, Eq. (4)–(5); §4.2/Appendix A.3] The defining contribution is the auxiliary future-frame loss L_vis, described as shaping the representation and action prior during training. This is what separates AeroAct from a video-backbone imitation policy. However, all reported ablations vary the reference-frame count or real-world fine-tuning; none varies λ_vis or trains with λ_vis=0. Tables 3–4 therefore cannot support the claim that future-frame prediction contributes to closed-loop performance. Please add a λ_vis ablation, including λ_vis=0, on both tracking and searching, with the same closed-loop metrics. Without this, the 'first WAM flight' claim has no empirical basis.
  2. [§4.2, Tables 3–4] The central temporal-context result is based on only 20 episodes per condition, with no error bars, multiple seeds, or confidence intervals. The differences between 1-frame and 9-frame conditions are large (e.g., 20% vs. 100% view success in Table 3), but at 20 episodes a single episode changes a percentage by 5 points, so distinctions such as AeroAct-FT 9-frame 95% vs. 100% are not separable. Moreover, no comparison is made to existing aerial VLN/VLA methods or to a conventional action-diffusion policy; the only baseline is the same model with fewer reference frames. Please report binomial confidence intervals or repeated seeds, and add at least one prior-method comparison.
  3. [§4.3] The real-world section reports a single flight with one instruction and no quantitative metric, repeated trials, or failure cases. This can demonstrate that the learned action interface can be executed on hardware, but it does not substantiate the 'first WAM-based real-world quadrotor flight' claim beyond a feasibility proof. Please add repeated runs with final-distance and success metrics, or explicitly narrow the claim to a feasibility demonstration. The paper's own Limitations section appropriately notes the short indoor scope, but it does not acknowledge that the WAM-specific loss component is untested.
minor comments (5)
  1. [§3.4, Eq. (6)] The self-guidance update is written in a way that is hard to follow: δt is defined as σ_i−σ_{i−1}, but the formula then mixes denoising-step notation with interval indices. Clarify the sampler step, the meaning of the prefix substitution, and why the second term uses a denoising step. The name 'self-guidance' may also be confused with classifier-free guidance; consider a more descriptive term.
  2. [§4.1] The fine-tuning protocol for AeroAct-FT is not specified. State how many real demonstrations are used, the learning rate and number of steps for fine-tuning, and which modules are frozen or updated.
  3. [Tables 1, 3, 4] Tables 1, 3, and 4 overlap heavily. Consolidate them or clearly cross-reference to avoid duplication and potential confusion about which numbers are the primary results.
  4. [§3.1, Eq. (2)–(3)] The same symbol g_Θ is used for both action prediction and future-frame prediction, with c_t introduced as an action latent. Clarify how c_t is obtained at inference and how it is combined with the context, since this is central to the action-centered design.
  5. [Appendix A.2] The sentence 'In the current logging code, dist_m and final dist_m denote the same final-distance value' reads like a leftover implementation note. Please remove it or integrate the notation cleanly, since it raises unnecessary doubts about the evaluation code.

Circularity Check

0 steps flagged

No circularity found: action and future-frame losses are separate, inference does not use imagined frames, and the temporal-context ablation is a genuine input change. Self-citations are infrastructure, not load-bearing.

full rationale

I walked the derivation chain. Equations (2)-(3) define action and future-frame predictions as separate outputs of the same model; Eq. (4)-(5) sum two losses over different targets (action noise epsilon_a vs. future-frame noise epsilon_f). The blockwise causal mask prevents action tokens from attending to future-frame tokens, and deployment omits T_f and uses only g_a, so no closed-loop result is equivalent to the future-frame supervision by construction. The temporal-context ablation (Tables 1-4, Appendix A.3) varies the number of reference frames h (1 vs. 9), which is a real input change, not a refit of the reported success/collision metrics. Self-guidance (Eq. 6a-6b) explicitly overwrites the overlapping prefix with the previous chunk's suffix; the paper presents this as a sampler constraint, not as a learned prediction, so it is not a hidden equation-level circularity. The self-citations (DiffAero [11], MAD [7]) supply simulation/dynamics infrastructure and contextual world-model examples; the central claims -- first physical WAM flight and 100% closed-loop success at 9 reference frames -- do not reduce to those citations. The WAM definition is attributed to prior work [9,10], not to a uniqueness theorem from this author group. The main substantive concern is the missing lambda_vis ablation (Eq. 4-5; no such sweep in Sec. 4.2 or A.3), meaning the contribution of L_vis is empirically unverified; if lambda_vis=0 performed identically, AeroAct would be a video-backbone imitation policy. That is an evidence gap and a correctness risk, not a circular reduction: no equation makes the action output a refitted version of the future-frame target. I therefore find no significant circularity, with only minor non-load-bearing self-citations.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

The central result relies on a substantial stack of unstated premises: transferred video priors, the untested visual-loss benefit, the fidelity of DiffAero dynamics, and the validity of handheld data as a drone-flight proxy. The main design parameters (reference-frame window, chunk horizon, loss weight, success thresholds) are chosen by hand and not explored.

free parameters (4)
  • Visual loss weight λ_vis = not reported
    Balances action loss and future-frame visual loss in Eq. (4); the paper never states its value or ablates it, yet it defines the WAM training objective.
  • Reference-frame count h and temporal stride Δ = h=9, Δ=3
    The central ablation (1 vs 9 reference frames) varies h; these values are chosen by hand and affect the temporal-context conclusion.
  • Trajectory segment duration T and chunk length p = T=2s, p=24
    Action representation parameters chosen by hand; success depends on this horizon covering the maneuvers needed for tracking and search.
  • Evaluation thresholds τ_c and τ_s = 0.25 and 2.0m
    Success thresholds are chosen by hand; changing them changes the reported success and collision rates.
axioms (6)
  • domain assumption Wan2.1-1.3B video diffusion transformer provides temporal/geometric priors that transfer to egocentric quadrotor vision
    Used as the frozen backbone in §3.2; the paper does not validate this transfer beyond its own results.
  • domain assumption Future-frame prediction loss L_vis improves the learned action distribution
    Core WAM premise from §3.2 Eq. (4)–(5); the paper reports no ablation removing L_vis.
  • domain assumption DiffAero dynamics and its flight controllers generate trajectories dynamically feasible and representative of real quadrotors
    All simulation data and the shared control interface in §3.3 rely on DiffAero [11].
  • domain assumption Handheld T265 odometry converted to quadrotor poses yields egocentric views valid for real-flight fine-tuning
    Real-world fine-tuning data in §3.3 depends on this proxy; no validation against actual drone footage is provided.
  • domain assumption Fifth-order polynomial endpoint parameterization can represent the maneuvers needed for tracking and object search
    The action space in Appendix A.1 restricts outputs to 2s endpoint segments; if this class is too narrow, the policy cannot express required behaviors.
  • standard math Diffusion/flow-matching objective with a causal mask trains a valid conditional action distribution
    Relies on standard diffusion/flow-matching theory [28]; no formal verification provided.

pith-pipeline@v1.3.0-alltime-deepseek · 11614 in / 16796 out tokens · 175786 ms · 2026-08-02T00:28:14.494682+00:00 · methodology

0 comments
read the original abstract

Language-conditioned quadrotor flight requires a policy to ground semantic goals, anticipate the visual consequences of ego-motion, and output control references that remain smooth and dynamically executable under rapidly changing first-person views. Existing aerial vision-language navigation and vision-language-action methods commonly use discrete actions, high-level waypoints, or instantaneous velocity commands, which provide limited supervision about how flight actions change future observations. We present AeroAct, an action-centered world-action model (WAM) for quadrotor navigation. To the best of our knowledge, AeroAct is the first WAM instantiated and demonstrated for real-world aerial flight. The model adapts a pretrained video diffusion Transformer to predict local trajectory-action chunks from egocentric visual history, proprioception, and language. Future first-person frames are used during training as dense consequence supervision, while deployment directly decodes actions without generating future video. To obtain aligned visual, state, language, and dynamically feasible action data, we build a DiffAero-based pipeline with complementary Isaac Lab and 3D Gaussian splatting renderers. We further introduce a low-cost handheld collection device that couples camera observations with motion estimates to recreate flight-like egocentric trajectories, and a self-guidance procedure that improves temporal consistency across overlapping trajectory chunks. Closed-loop simulation and real-world experiments show that temporal visual context improves target tracking and object-search performance, and that WAM-based policies can be executed on a physical quadrotor.

Figures

Figures reproduced from arXiv: 2607.14997 by Gang Wang, Haolin Chen, Jian Sun, Qiyuan Zhu, Runqing Wang, Xinhong Zhang, Xinjiang Wang, Yubo Huang, Yuhao Mo, Yu Hu, Zhongxin Chen.

Figure 1
Figure 1. Figure 1: Overview of AeroAct. The policy receives egocentric visual history, proprioception and language, and predicts a chunked local trajectory for aerial navigation. Training uses paired flight actions and future visual observations from Isaac, DiffAero, 3DGS scenes, and real-world handheld data; deployment decodes actions with self-guidance and executes these actions through a trajectory controller. Abstract: L… view at source ↗
Figure 2
Figure 2. Figure 2: Action-centered world-action Transformer. Visual history, proprioception, and lan￾guage condition the action decoder. Future-frame tokens are used during training as consequence supervision, but are masked from action tokens and can be disabled at inference time. an action-centered design for efficient control [9]. These methods are mainly developed for table￾top manipulation, where cameras are comparative… view at source ↗
Figure 3
Figure 3. Figure 3: Simulation data pipeline. DiffAero provides a shared dynamics and control interface. Isaac Lab improves visual fidelity, while the 3DGS branch scales scene and object diversity with lower asset-engineering cost. so we encode and forecast one observation every ∆ steps. At training time, a clip ot−(h−1)∆:t+K∆ is encoded by the frozen video VAE into spatiotemporal latent tokens. After 3D RoPE positional encod… view at source ↗
Figure 4
Figure 4. Figure 4: Inference-time self-guidance. Consecutive action chunks overlap in time. The prefix of the new chunk is guided by the suffix of the previous chunk, reducing discontinuities before the remaining horizon is sampled. motion anticipation and visual servoing. Second, in Isaac reaching, a language-specified object is placed among distractors and the quadrotor approaches it, linking language grounding to goal￾dir… view at source ↗
Figure 5
Figure 5. Figure 5: Experimental hardware platforms. (a) UAV platform for flight-task execution, equipped with onboard sensing modules and connected to a remote server via a local network for offboard inference. The resulting trajectories are tracked using OM-MPC. (b) Self-made portable data￾collection platform for recording cropped camera observations and corresponding trajectory an￾notations, enabling efficient large-scale … view at source ↗
Figure 6
Figure 6. Figure 6: Real-world flight with language-conditioned WAM. AeroAct generates feasible flight commands and guides the quadrotor toward the specified target object in a lab environment. The trajectory illustrates closed-loop execution of the learned WAM policy on a physical aerial robot. To the best of our knowledge, this is the first WAM-based real-world quadrotor flight demonstra￾tion, where a learned model directly… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

32 extracted references · 4 linked inside Pith

  1. [1]

    Hafner, J

    D. Hafner, J. Pasukonis, J. Ba, and T. Lillicrap. Mastering diverse control tasks through world models.Nature, 640(8059):647–653, Apr. 2025. ISSN 1476-4687. doi:10.1038/ s41586-025-08744-2

    2025

  2. [2]

    Zhang, G

    W. Zhang, G. Wang, J. Sun, Y . Yuan, and G. Huang. STORM: Efficient stochastic transformer based world models for reinforcement learning. InProc. Adv. Neural Inf. Process. Syst., 2023

    2023

  3. [3]

    Bruce, M

    J. Bruce, M. Dennis, A. Edwards, J. Parker-Holder, Y . Shi, E. Hughes, M. Lai, A. Mavalankar, R. Steigerwald, C. Apps, Y . Aytar, S. Bechtle, F. Behbahani, S. Chan, N. Heess, L. Gonzalez, S. Osindero, S. Ozair, S. Reed, J. Zhang, K. Zolna, J. Clune, N. de Freitas, S. Singh, and T. Rockt¨aschel. Genie: Generative interactive environments, 2024

    2024

  4. [4]

    S. Yang, Y . Du, S. K. S. Ghasemipour, J. Tompson, L. P. Kaelbling, D. Schuurmans, and P. Abbeel. Learning interactive real-world simulators. InThe Twelfth International Conference on Learning Representations, 2024

    2024

  5. [5]

    Romero, A

    A. Romero, A. Shenai, I. Geles, E. Aljalbout, and D. Scaramuzza. Dream to fly: Model-based reinforcement learning for vision-based drone flight.arXiv:2501.14377, 2025

    Pith/arXiv arXiv 2025

  6. [6]

    Zhang, B

    Y . Zhang, B. Gao, G. Wang, J. Sun, and Z. Li. CORB-Planner: Corridor as observations for rl planning in high-speed flight.IEEE/ASME Trans. Mechatron., 2025. doi:10.1109/TMECH. 2025.3620233

    arXiv 2025

  7. [7]

    Zhang, R

    X. Zhang, R. Wang, Y . Ren, D. Yu, B. Zhou, J. Sun, F. Deng, J. Chen, and G. Wang. Mad: Mapping-aware world models for agile quadrotor flight.arXiv preprint arXiv:2606.04534, 2026

    Pith/arXiv arXiv 2026

  8. [8]

    Z. Feng, R. Xue, L. Yuan, Y . Yu, N. Ding, M. Liu, B. Gao, J. Sun, X. Zheng, and G. Wang. Multi-agent embodied ai: Advances and future directions.Science China Information Sci- ences, 69(5):151202, 2026

    2026

  9. [9]

    A. Ye, B. Wang, C. Ni, G. Huang, G. Zhao, H. Li, H. Li, J. Li, J. Lv, J. Liu, M. Cao, P. Li, Q. Deng, W. Mei, X. Wang, X. Chen, X. Zhou, Y . Wang, Y . Chang, Y . Li, Y . Zhou, Y . Ye, Z. Liu, and Z. Zhu. GigaWorld-Policy: An efficient action-centered world–action model, 2026

    2026

  10. [10]

    S. Ye, Y . Ge, K. Zheng, S. Gao, S. Yu, G. Kurian, S. Indupuru, Y . L. Tan, C. Zhu, J. Xi- ang, A. Malik, K. Lee, W. Liang, N. Ranawaka, J. Gu, Y . Xu, G. Wang, F. Hu, A. Narayan, J. Bjorck, J. Wang, G. Kim, D. Niu, R. Zheng, Y . Xie, J. Wu, Q. Wang, R. Julian, D. Xu, Y . Du, Y . Chebotar, S. Reed, J. Kautz, Y . Zhu, L. J. Fan, and J. Jang. World action m...

    2026

  11. [11]

    Zhang, R

    X. Zhang, R. Wang, Y . Ren, J. Sun, H. Fang, J. Chen, and G. Wang. DiffAero: A GPU- accelerated differentiable simulation framework for efficient quadrotor policy learning, 2025. 9

    2025

  12. [12]

    Kerbl, G

    B. Kerbl, G. Kopanas, T. Leimk ¨uhler, and G. Drettakis. 3d gaussian splatting for real-time radiance field rendering.ACM Transactions on Graphics, 42(4):139:1–139:14, 2023. doi: 10.1145/3592433

    doi:10.1145/3592433 2023

  13. [13]

    Y . Jia, G. Wang, Y . Dong, J. Wu, Y . Zeng, H. Lin, Z. Wang, H. Ge, W. Gu, C. Li, Z. Wang, Y . Cheng, W. Sui, R. Huang, and G. Zhou. Discoverse: Efficient robot simulation in complex high-fidelity environments.arXiv preprint arXiv:2507.21981, 2025

    Pith/arXiv arXiv 2025

  14. [14]

    Y . Jia, H. Zhang, Z. Zhang, J. Wu, M. Yu, Z. Wang, D. Jiang, Z. Li, C. Cao, Z. Yu, X. Yang, H. Ge, Y . Zhang, J. Zhang, Z. Huang, T. Liu, S. Chen, J. Wang, B. Xie, X. Yao, X. Deng, G. Wang, J. Zhang, L. Hao, Z. Chen, Y . Chen, A. Wang, H. Tian, Y . Yan, Z. Cao, Y . Jiang, H. Shao, Y . Li, L. Shi, B. Chen, W. Sui, H. Cui, Y . Qin, R. Huang, L. Han, T. Wan...

    Pith/arXiv arXiv 2026

  15. [15]

    Y . Fan, W. Chen, T. Jiang, C. Zhou, Y . Zhang, and X. E. Wang. Aerial vision-and-dialog navigation, 2023

    2023

  16. [16]

    S. Liu, H. Zhang, Y . Qi, P. Wang, Y . Zhang, and Q. Wu. AerialVLN: Vision-and-language navigation for uavs, 2023

    2023

  17. [17]

    X. Wang, D. Yang, Z. Wang, H. Kwan, J. Chen, W. Wu, H. Li, Y . Liao, and S. Liu. Towards realistic UA V vision-language navigation: Platform, benchmark, and methodology. InThe Thirteenth International Conference on Learning Representations, 2025

    2025

  18. [18]

    Y . Gao, C. Li, Z. You, J. Liu, L. Zhen, P. CHEN, Q. Chen, Z. Tang, L. Wang, Yangpenghui, Y . Tang, Y . Tang, S. Liang, S. Zhu, Z. Xiong, Y . Su, X. Ye, J. Li, Y . Ding, D. Wang, Z. Wang, B. Zhao, and X. Li. OpenFly: A comprehensive platform for aerial vision-language navigation. InThe F ourteenth International Conference on Learning Representations, 2026

    2026

  19. [19]

    X. Wang, D. Yang, Y . Liao, W. Zheng, wenjun wu, B. Dai, H. Li, and S. Liu. UA V- Flow Colosseo: A real-world benchmark for flying-on-a-word UA V imitation learning. In The Thirty-ninth Annual Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2026

    2026

  20. [20]

    Zheng, Y

    G. Zheng, Y . Ban, M. Zhang, J. Zheng, and B. Zhou. OnFly: Onboard zero-shot aerial vision- language navigation toward safety and efficiency, 2026

    2026

  21. [21]

    Y . Wu, M. Zhu, X. Li, Y . Du, Y . Fan, W. Li, Z. Han, X. Zhou, and F. Gao. VLA-AN: An efficient and onboard vision-language-action framework for aerial navigation in complex envi- ronments, 2025

    2025

  22. [22]

    C. Zhu, R. Yu, S. Feng, B. Burchfiel, P. Shah, and A. Gupta. Unified World Models: Coupling video and action diffusion for pretraining on large robotic datasets. InProceedings of Robotics: Science and Systems (RSS), 2025

    2025

  23. [23]

    M. J. Kim, Y . Gao, T.-Y . Lin, Y .-C. Lin, Y . Ge, G. Lam, P. Liang, S. Song, M.-Y . Liu, C. Finn, and J. Gu. Cosmos Policy: Fine-tuning video models for visuomotor control and planning, 2026

    2026

  24. [24]

    H. Bi, H. Tan, S. Xie, Z. Wang, S. Huang, H. Liu, R. Zhao, Y . Feng, C. Xiang, Y . Rong, H. Zhao, H. Liu, Z. Su, L. Ma, H. Su, and J. Zhu. Motus: A unified latent action world model, 2025

    2025

  25. [25]

    T. Yuan, Z. Dong, Y . Liu, and H. Zhao. Fast-W AM: Do world action models need test-time future imagination?, 2026. 10

    2026

  26. [26]

    T. Wan, A. Wang, B. Ai, B. Wen, C. Mao, C.-W. Xie, D. Chen, F. Yu, H. Zhao, J. Yang, J. Zeng, J. Wang, J. Zhang, J. Zhou, J. Wang, J. Chen, K. Zhu, K. Zhao, K. Yan, L. Huang, M. Feng, N. Zhang, P. Li, P. Wu, R. Chu, R. Feng, S. Zhang, S. Sun, T. Fang, T. Wang, T. Gui, T. Weng, T. Shen, W. Lin, W. Wang, W. Wang, W. Zhou, W. Wang, W. Shen, W. Yu, X. Shi, X....

    2025

  27. [27]

    HY-World, C

    T. HY-World, C. Cao, X. Zuo, Z. Wang, Y . Zhang, J. Wu, Z. Liu, Y . Gong, Y . Liu, B. Yuan, C. Zhang, C. Li, D. Guo, F. Yang, H. Zhang, H. Cao, J. Zhu, J. Lin, J. Xiao, J. Zhang, J. Yu, L. Wang, L. Wang, L. Wang, Linus, M. Chen, P. He, P. Zhao, Q. Chen, R. Chen, R. Shao, S. Liu, W. Qin, X. Niu, X. Yuan, Y . Sun, Y . Tang, Y . Sun, Y . Lian, Y . Tan, Y . L...

    2026

  28. [28]

    Lipman, R

    Y . Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le. Flow matching for generative modeling. InThe Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023

    2023

  29. [29]

    Loshchilov and F

    I. Loshchilov and F. Hutter. Decoupled weight decay regularization. In7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net, 2019

    2019

  30. [30]

    G. Lu, W. Xu, and F. Zhang. On-Manifold model predictive control for trajectory tracking on robotic systems.IEEE Transactions on Industrial Electronics, 70(9):9192–9202, 2023. doi:10.1109/TIE.2022.3212397

    arXiv 2023

  31. [31]

    J. Lu, X. Zhang, H. Shen, L. Xu, and B. Tian. You only plan once: A learning-based one-stage planner with guidance learning.IEEE Robotics and Automation Letters, 9(7):6083–6090, July

  32. [2024]

    11 A Appendix A.1 Observation and Action Spaces Proprioception and actions are expressed in a yaw-aligned local frame

    doi:10.1109/LRA.2024.3399589. 11 A Appendix A.1 Observation and Action Spaces Proprioception and actions are expressed in a yaw-aligned local frame. Itsx-axis lies in the vertical plane defined by the bodyx-axis, and itsz-axis points upward. The proprioceptive input is s= [v l⊤,a l⊤,u l⊤ z ]⊤ ∈R 9,(7) wherev l,a l, andu l z are the local velocity, local a...

    arXiv 2024