Pith. sign in

REVIEW 5 major objections 6 minor 29 references

A reinforcement-learning agent can perform a dolly-in cinematic shot at the same level as a finely hand-tuned PD controller, and its policy transfers to a real robot without retraining.

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 →

TD3-based RL matches a hand-tuned PD controller for dolly-in shots on a ground robot and transfers to real hardware with moderate sim-to-real correlation.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A genuine but modest TD3 application to ground-based dolly-in film shots; the abstract and sim-to-real correlation claims outrun the evidence. the 5 major comments →

arxiv 2509.00564 v1 pith:6JUORHSX submitted 2025-08-30 cs.RO cs.CVcs.LG

Reinforcement Learning of Dolly-In Filming Using a Ground-Based Robot

classification cs.RO cs.CVcs.LG
keywords reinforcement learningdolly-in shotrobotic cinematographysim-to-real transferTD3ground-based robotcamera control
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 reading

This paper asks whether a ground-based filming robot can learn a classic cinematic move—the dolly-in, where the camera moves smoothly toward a subject while keeping it centred and growing in the frame—without an explicit model of the robot or camera dynamics. The authors train a TD3 reinforcement-learning agent on four simultaneous actions (throttle, steering, camera pan, and camera tilt) using rewards defined directly on image-derived quantities: subject area, centroid position, and offset angle. Their central claim is that this combined-control agent matches a finely hand-tuned PD controller in simulation, and that the same policy, applied with no fine-tuning to a real modified ROSBot 2.0, preserves performance with a positive simulation-to-real correlation. They also show that a single combined controller outperforms two independent throttle and steering agents in task reward, though the independent agents converge faster. If true, the result matters because it shows standard continuous-control RL can replace hand-tuned classical controllers on cheap ground hardware and take over repetitive camera movements.

Core claim

On the paper's own terms, the central discovery is that an RL agent using its full range of motion—throttle, steering, pan, and tilt—reaches dolly-in performance comparable to a finely hand-tuned PD controller in simulation, measured with a normalized error metric on subject area and position across 100 trials. The same trained policy also executes the shot zero-shot on a physical robot built from a ROSBot 2.0 with a camera turret; across 30 real runs from three start poses, the simulation-to-real correlation coefficient (SRCC) is positive, roughly 0.46–0.83 depending on metric and pose. The paper further establishes that a single combined throttle-and-steering agent achieves better mean rew

What carries the argument

The paper's load-bearing machinery is a reward system built from image moments plus a heading offset term. The subject's size is the zeroth geometric moment of a colour-thresholded mask; its position is the mask centroid; an offset angle combines the pixel-to-centre camera angle with the pan-joint angle. Each reward component is a normalized relative error between current and target values, so the agent is pushed toward a shot in which the subject occupies a set fraction of the frame, sits at frame centre, and lies directly ahead of the moving robot. A TD3 agent—a standard off-policy actor-critic algorithm for continuous control—maps nine state variables (area, area error, centroid errors, p

Load-bearing premise

The transfer claim rests on the assumption that the physics simulation is a faithful digital twin of the real robot's dynamics, camera optics, and ground interaction; the paper reports no direct comparison of simulated versus real trajectories, only 30 real runs and run-level reward correlations, so if the simulator diverges from the real robot in any of those details the zero-shot result does not establish a general capability.

What would settle it

Run the trained policy from the same three start poses in simulation and on the physical robot, and record per-timestep distance to subject, heading error, and pan/tilt angles. If the real robot's heading error systematically diverges from simulation (for instance, simulated error stays under 3° while real error crosses 8°) even though cumulative rewards stay similar, the digital-twin premise is falsified and the SRCC claim is not about policy fidelity. A cheaper probe: perturb simulator wheel friction or camera field of view by ±10% and see whether area and centroid targets are still held; a

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

If this is right

  • The same TD3 pipeline can be re-run for other scripted camera moves—push-ins at different speeds, subject tracking, arc shots—by changing only the target values in the reward functions, not the learning algorithm.
  • Combined single-agent control should be preferred over independent throttle and steering agents for multi-action filming tasks when coordination matters, accepting a longer training time for better task reward.
  • Zero-shot sim-to-real transfer is feasible for ground-based camera robots when the simulator mirrors the real platform's dynamics and optics; the paper's positive SRCC values are evidence that such transfer works for this task.
  • A classic PD controller remains a strong baseline, but RL removes the need to hand-tune it per shot or per platform, since the policy is learned from reward rather than from gains.
  • The framewise visual state abstraction makes the formulation platform-agnostic, so the same reward and state space could transfer to other robot bodies with retraining in a matching simulator.

Where Pith is reading between the lines

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

  • I infer that the same camera-centred state abstraction—subject area, frame-centroid error, and offset angle—could be lifted onto other camera carriers, such as a gimbal or drone, by re-simulating only the carrier's dynamics, since the reward is written on the frame rather than on the robot's kinematics.
  • Because the reward weights are explicit parameters, the pipeline could become a parameterised shot language: a director could specify 'slow approach', 'tighter final framing', or 'subject slightly off-centre' as weight vectors without redesigning the reward.
  • A natural next measurement the paper does not report is per-timestep agreement between simulated and real state trajectories (distance, heading, pan, tilt); the reported SRCC is computed from run-level rewards, so I would expect that stricter trajectory metric to be strong during the steady approach and weaker in the first few timesteps of alignment.
Share X Bluesky LinkedIn Reddit HN

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

5 major / 6 minor

Summary. The paper applies Twin Delayed Deep Deterministic Policy Gradient (TD3) to automate dolly-in camera shots on a ground-based robot, comparing independent throttle/steering agents against a combined agent and a 'complex' agent that also controls pan and tilt. The authors report that the combined agent outperforms independent control, that the complex RL agent achieves performance comparable to a hand-tuned PD controller in simulation, and that the policy transfers zero-shot to a real modified ROSBot 2.0, with Spearman Rank Correlation Coefficient (SRCC) values claimed to indicate strong simulation-to-reality correlation. The paper also introduces task formalization and reward shaping based on subject area, position, and offset angle.

Significance. If substantiated, the paper would provide a useful contribution to robotic cinematography by showing that a standard RL algorithm can learn a scripted camera maneuver on a ground platform without explicit dynamics modeling, and can transfer to real hardware. The comparative evaluation of combined versus independent control is also of practical interest. However, the central evidence is weakened by reward/evaluation circularity, an overclaimed abstract, statistically fragile sim-to-real correlation measures, and an unvalidated digital-twin assumption. The paper does not provide open-source code or detailed hyperparameter tables, limiting reproducibility. The strengths are the clear task formalization and the attempt to quantify sim-to-real transfer, but these strengths are undermined by the statistical and methodological issues below.

major comments (5)
  1. [Abstract and Section IV-C-2] The abstract claims the RL pipeline 'surpasses traditional Proportional-Derivative controller performance in simulation', but Section IV-C-2 and Figure 5b only support 'comparable' performance. Moreover, the PD controller is described as 'hand-tuned to achieve a performance goal aligned with that of the complex agent' (Section IV-C-2); this tuning protocol biases the comparison and does not support a 'surpasses' claim. The manuscript should either present statistical evidence of superiority or temper the abstract to match the body.
  2. [Section IV-D, Table IV] The zero-shot sim-to-real transfer claim rests on SRCC values computed from only 10 runs per starting position, with no p-values, confidence intervals, or correction for multiple comparisons. As defined by Kadian et al. [29], SRCC is intended to measure correlation across a set of conditions (e.g., scenes); here, only three starting positions are used, so if SRCC is computed per-position over 10 runs, it is not the cross-condition measure, and if computed across the three positions, n=3 yields meaningless rank correlations. Values like 0.46 and 0.56 are not conventionally 'strong'. The paper must either recompute SRCC with a valid sample size and report significance, or provide a different quantitative comparison (e.g., trajectory errors) to support contribution 3.
  3. [Section III-A-3 and Section IV-C-2] The reward functions (Eqs. 7-10) are built directly from the same quantities used for evaluation: subject area, centroid position, and offset angle. Consequently, reporting higher reward or lower normalized error amounts to measuring optimization of the reward itself, not an independently defined cinematic quality. This is acceptable in RL practice, but the paper should acknowledge that the claimed 'performance comparable to PD' is partly a statement about reward optimization, and should include at least one metric not used in the reward (e.g., smoothness of the camera trajectory, final framing error under disturbance) to avoid circularity.
  4. [Section IV-D and Section V] The sim-to-real transfer experiment assumes the PyBullet simulation is a high-fidelity digital twin of the real ROSBot 2.0, but the paper provides no comparison of simulated versus real trajectories, no domain randomization, and only 30 real-world runs. The Discussion even states that 'the need for high-fidelity digital twins becomes evident' (Section V), which is a concession that the current simulation is not validated. Without such validation, the zero-shot transfer claim is fragile. The authors should either provide evidence of trajectory/physics fidelity or explicitly limit the claim to the specific test conditions.
  5. [Section IV-C-1, Figure 4] The independent-versus-combined experiment reports mean rewards of -120 versus -160 over 100 trials, but no standard deviations, box plots, or significance tests are given. The claim that combined control 'outperformed' independent control is therefore not statistically supported. Since this is the first contribution, the authors should report variance and perform a paired test across the 100 trials.
minor comments (6)
  1. [Section III-A-3, Eq. (9)] The definition of Rarea,scaled uses 'p' and 'k' without specifying that p is the current area and k is the threshold. Also, the formula for Rarea,scaled is incomplete; it references Rarea,original but the scaled version's connection to Eq. (6) is not clearly derived.
  2. [Section III-B] The hyperparameter list omits the reward weights w1, w2, w3 and the action penalty magnitude p in Eq. (10). These are free parameters that affect the reported results and should be reported.
  3. [Section IV-C-2] The PD controller gains are not reported, making it impossible to assess the 'finely hand-tuned' claim. Include gains and the tuning procedure.
  4. [Section IV-B] Figure 2b is a useful diagram, but the text says 'an external computer powered by an Intel(R) Core(TM) i7-10875H CPU and Nvidia RTX 2080 Super GPU supports image processing and RL computations' — clarify how RL is executed on the real robot (on-device or via ROS over network) and whether latency is accounted for.
  5. [General] Several equations and symbols are not formally defined (e.g., 's' in Eq. (2), 'pE' and 'pmax' in Eq. (8), 'pA' and 'pE' in Eq. (9) are reused). A notation table would help.
  6. [References] Reference [17] is cited as 'Opencv 4.5.0' but the author list is incomplete; also reference [26] is a GitHub repository without a version or commit hash.

Circularity Check

0 steps flagged

No significant circularity: the reward design and evaluation share the same objective, but the central claims are empirical and externally tested via real-world transfer.

full rationale

The paper's derivation chain is not circular. The dolly-in shot is formally defined in Section II via subject area, centroid position, and offset angle; these same quantities are then used to shape the reward functions in Section III-A (Eqs. 7–10). This overlap is standard reinforcement learning practice: the reward encodes the task objective, and evaluating the agent on that same objective measures how well it optimized the given reward. The central claim—that a TD3 agent can learn a policy that achieves this objective in simulation and transfer it zero-shot to a real robot—is an empirical result, not a tautology. The real-world experiments in Section IV-D provide external evidence independent of the training reward. The PD controller baseline was hand-tuned to the same task goal, not to the RL agent's achieved performance, so the comparison is not constructed to be equal. There are no load-bearing self-citations, no imported uniqueness theorems, and no renamed known results presented as new derivations. The concerns about SRCC statistical significance and digital-twin fidelity are validity/robustness issues, not circularity. The paper does not reduce any prediction to its own inputs by construction.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The paper's results depend on several hand-chosen reward and baseline parameters (reward weights, threshold k, action penalty, PD gains) that are not fully reported. The central claims also rest on domain assumptions: that the dolly-in is fully captured by the three geometric metrics, that the simulation is a faithful digital twin, and that color thresholding reliably segments the subject. No new physical entities are introduced.

free parameters (4)
  • Reward weights w1, w2, w3 = not reported
    Used in Eqs. 9 and 10 to balance area, position, and offset objectives; chosen by hand and never reported numerically, so the trained policy is not uniquely specified.
  • Threshold k in Rarea_scaled = not reported
    Defines the crossover in the piecewise area reward (Eq. 9); its value affects the policy and is not stated.
  • Action penalty magnitude p in Rcomplex = not reported
    Penalty for large action changes between timesteps; threshold and weight not specified.
  • PD controller gains = hand-tuned
    The comparison baseline is tuned 'to achieve a performance goal aligned with that of the complex agent', so the gains are fitted to match the RL agent, biasing the comparison.
axioms (4)
  • domain assumption The dolly-in shot is fully specified by subject area, position, and offset angle (Section II).
    The reward and all evaluation metrics are built from these three quantities. If cinematic quality depends on other factors, such as velocity profile or background composition, the claim of achieving dolly-in shots is not fully captured.
  • domain assumption The PyBullet simulation is a faithful digital twin of the real robot (Sections IV-B and IV-D).
    Zero-shot transfer assumes the sim dynamics, camera model, and friction match the real platform closely enough; no validation of this fidelity is provided.
  • standard math TD3 converges to a good policy given the chosen hyperparameters and reward; standard RL convergence assumptions (Sutton and Barto, Fujimoto et al.)
    The paper relies on TD3's properties but does not analyze convergence for this specific problem; results are empirical.
  • domain assumption Colour thresholding reliably segments the subject in both simulation and real world (Section II-A).
    The authors acknowledge 'the simplistic subject detection method required stable and consistent conditions', which limits the real-world validity.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Reinforcement Learning of Dolly-In Filming Using a Ground-Based Robot." pith.science (2026). https://pith.science/paper/6JUORHSX

@misc{pith2026250900564,
  author       = {Pith},
  title        = {Pith review of: Reinforcement Learning of Dolly-In Filming Using a Ground-Based Robot},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6JUORHSX}},
  note         = {Machine review of arXiv:2509.00564}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Free-roaming dollies enhance filmmaking with dynamic movement, but challenges in automated camera control remain unresolved. Our study advances this field by applying Reinforcement Learning (RL) to automate dolly-in shots using free-roaming ground-based filming robots, overcoming traditional control hurdles. We demonstrate the effectiveness of combined control for precise film tasks by comparing it to independent control strategies. Our robust RL pipeline surpasses traditional Proportional-Derivative controller performance in simulation and proves its efficacy in real-world tests on a modified ROSBot 2.0 platform equipped with a camera turret. This validates our approach's practicality and sets the stage for further research in complex filming scenarios, contributing significantly to the fusion of technology with cinematic creativity. This work presents a leap forward in the field and opens new avenues for research and development, effectively bridging the gap between technological advancement and creative filmmaking.

Figures

Figures reproduced from arXiv: 2509.00564 by Alan Hunter, Jack Saunders, Philip Lorimer, Wenbin Li.

Figure 1
Figure 1. Figure 1: Sequence of the dolly-in shot in the real world, showing the RL agent guiding our ground-based filming robot during [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Modified ROSBot 2.0 for Filming: Physical Modifi [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Training and stability evaluation for each agent. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of Combined and Independent control [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Training and evaluation for complex agent. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] 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

29 extracted references · 27 canonical work pages · 1 internal anchor

  1. [1]

    The role of robots in entertainment,

    R. Bogue, “The role of robots in entertainment,” Industrial robot an international journal., vol. 49, no. 4, pp. 667–671, 2022

  2. [2]

    R. S. Sutton and A. G. Barto, Reinforcement learning: An introduction. MIT press, 2018

  3. [3]

    Aerial filming with synchronized drones using reinforcement learning,

    K. C. W. Goh, R. B. C. Ng, Y . K. Wong, N. J. H. Ho, and M. C. H. Chua, “Aerial filming with synchronized drones using reinforcement learning,” Multimedia Tools and Applications , vol. 80, pp. 18125 – 18150, 2021

  4. [4]

    Autonomous camera systems: A survey,

    J. Chen and P. Carr, “Autonomous camera systems: A survey,” AAAI Workshop - Technical Report, vol. WS-14-06, pp. 18–22, 2014

  5. [5]

    Hybrid robotic/virtual pan-tilt- zom cameras for autonomous event recording,

    P. Carr, M. Mistry, and I. Matthews, “Hybrid robotic/virtual pan-tilt- zom cameras for autonomous event recording,” in Proceedings of the 21st ACM International Conference on Multimedia , MM ’13, (New York, NY , USA), p. 193–202, Association for Computing Machinery, 2013

  6. [6]

    Virtual forces for camera planning in smart vision systems,

    A. Farag and A. Abdel-Hakim, “Virtual forces for camera planning in smart vision systems,” pp. 269–274, 02 2005

  7. [7]

    Computational sports broadcasting: Automated director assistance for live sports,

    C. Chen, O. Wang, S. Heinzle, P. Carr, A. Smolic, and M. Gross, “Computational sports broadcasting: Automated director assistance for live sports,” Proceedings - IEEE International Conference on Multimedia and Expo , 2013

  8. [8]

    Reinforcement learning in robotic applications: A comprehensive survey,

    B. Singh, R. Kumar, and V . P. Singh, “Reinforcement learning in robotic applications: A comprehensive survey,” Artif. Intell. Rev. , vol. 55, p. 945–990, feb 2022

  9. [9]

    Reinforcement learning for robot research: A comprehensive review and open issues,

    T. Zhang and H. Mo, “Reinforcement learning for robot research: A comprehensive review and open issues,” International Journal of Advanced Robotic Systems , vol. 18, no. 3, p. 17298814211007305, 2021

  10. [10]

    Can a Robot Become a Movie Director? Learning Artistic Principles for Aerial Cinematography

    M. Gschwindt, E. Camci, R. Bonatti, W. Wang, E. Kayacan, and S. Scherer, “Can a Robot Become a Movie Director? Learn- ing Artistic Principles for Aerial Cinematography,” arXiv e-prints , p. arXiv:1904.02579, Apr. 2019

  11. [11]

    Deep reinforcement learning for controlling frontal person close-up shooting,

    N. Passalis and A. Tefas, “Deep reinforcement learning for controlling frontal person close-up shooting,” vol. 335, pp. 37–47

  12. [12]

    Decentralized reinforcement learning applied to mobile robots,

    D. L. Leottau, A. Vatsyayan, J. R. del Solar, and R. Babu ˇska, “Decentralized reinforcement learning applied to mobile robots,” in Robot Soccer World Cup , 2016

  13. [13]

    Fractional-order active disturbance rejection control with fuzzy self-tuning for precision stabilized plat- form,

    J. Zhao, T. Zhao, and N. Liu, “Fractional-order active disturbance rejection control with fuzzy self-tuning for precision stabilized plat- form,” Entropy, vol. 24, no. 11, 2022

  14. [14]

    A practical guide to multi-objective reinforcement learning and planning,

    C. F. Hayes, R. R ˘adulescu, E. Bargiacchi, J. K¨allstr¨om, M. Macfarlane, M. Reymond, T. Verstraeten, L. M. Zintgraf, R. Dazeley, F. Heintz, E. Howley, A. A. Irissappane, P. Mannion, A. Now ´e, G. Ramos, M. Restelli, P. Vamplew, and D. M. Roijers, “A practical guide to multi-objective reinforcement learning and planning,” Autonomous Agents and Multi-Agen...

  15. [15]

    Glossary of Virtual Reality Terminol- ogy,

    C. Manetta and R. A. Blade, “Glossary of Virtual Reality Terminol- ogy,” International Journal of Virtual Reality, vol. 1, no. 2, pp. 35–39, 1995

  16. [16]

    Static saliency vs. Dynamic saliency: A comparative study,

    T. V . Nguyen, M. Xu, G. Gao, M. Kankanhalli, Q. Tian, and S. Yan, “Static saliency vs. Dynamic saliency: A comparative study,” MM 2013 - Proceedings of the 2013 ACM Multimedia Conference, pp. 987– 996, 2013

  17. [17]

    Opencv 4.5.0,

    G. Bradski and A. Kaehler, “Opencv 4.5.0,” 2023

  18. [18]

    Image analysis using moments,

    L. G. Kotoulas and I. Andreadis, “Image analysis using moments,” 2005

  19. [19]

    Reward function and initial values: Better choices for accelerated goal-directed rein- forcement learning,

    L. Matignon, G. J. Laurent, and N. Le Fort-Piat, “Reward function and initial values: Better choices for accelerated goal-directed rein- forcement learning,” Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 4131 LNCS, pp. 840–849, 2006

  20. [20]

    Addressing function approx- imation error in actor-critic methods,

    S. Fujimoto, H. Hoof, and D. Meger, “Addressing function approx- imation error in actor-critic methods,” in Proceedings of the 35th International Conference on Machine Learning , vol. 80, 2018

  21. [21]

    Predicting optimal value functions by inter- polating reward functions in scalarized multi-objective reinforcement learning,

    A. Kusari and J. P. How, “Predicting optimal value functions by inter- polating reward functions in scalarized multi-objective reinforcement learning,” pp. 7484–7490. ISBN: 9781728173955

  22. [22]

    Pybullet, a python module for physics sim- ulation for games, robotics and machine learning

    E. Coumans and Y . Bai, “Pybullet, a python module for physics sim- ulation for games, robotics and machine learning.” http://pybullet.org, 2016–2023

  23. [23]

    Pytorch: An imperative style, high- performance deep learning library,

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Kopf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala, “Pytorch: An imperative style, high- performance deep learning library,” in Advances in Neural Information Processing ...

  24. [24]

    Openai gym,

    G. Brockman, V . Cheung, L. Pettersson, J. Schneider, J. Schul- man, J. Tang, and W. Zaremba, “Openai gym,” arXiv preprint arXiv:1606.01540, 2016

  25. [25]

    Rosbot 2.0 official documentation

    Husarion, “Rosbot 2.0 official documentation.” https://husarion.com/ manuals/rosbot-manual/, 2020

  26. [26]

    Interbotix ros turrets

    Interbotix, “Interbotix ros turrets.” https://github.com/Interbotix/ interbotix ros turrets/tree/main/interbotix ros xsturrets, 2022. GitHub repository

  27. [27]

    Robotic operating system

    Stanford Artificial Intelligence Laboratory et al., “Robotic operating system.”

  28. [28]

    Frech, simple-pid: A Simple PID Controller API for Python , 2023

    J. Frech, simple-pid: A Simple PID Controller API for Python , 2023

  29. [29]

    Sim2real predictivity: Does evaluation in simulation predict real-world performance,

    A. Kadian, J. Truong, A. Gokaslan, A. Clegg, E. Wijmans, S. Lee, M. Savva, S. Chernova, and D. Batra, “Sim2real predictivity: Does evaluation in simulation predict real-world performance,” IEEE Robotics and Automation Letters , vol. PP, pp. 1–1, 08 2020

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.