Pith. sign in

REVIEW 4 major objections 5 minor 23 references

A humanoid robot can learn complex skills from text-prompt videos alone, with no real-world demonstration data.

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 ↗

Text-prompt-generated synthetic videos can serve as reference demonstrations for teaching a simulated humanoid robot several whole-body tasks.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection A plausible synthetic-video-to-humanoid pipeline that is demonstrated only in simulation and lacks the baselines and feasibility checks needed to support its core claim. the 4 major comments →

arxiv 2607.21648 v1 pith:JIPLXROS submitted 2026-07-22 cs.RO

Learning Diverse Humanoid Tasks via Synthetic Video Scenarios without Real World Data

Yun-Hao Tsai , Cong-Thanh Vu , Yen-Chen Liu This is my paper
classification cs.RO
keywords synthetic demonstrationshumanoid robotreinforcement learningtext-to-video generationmotion retargetingimitation learningmotion chainingsimulation
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

The paper claims that reference motions extracted from AI-generated videos can replace motion-capture data for imitation-based humanoid reinforcement learning. By writing a text prompt, the framework generates multiple videos of the same task, converts them into robot-feasible joint trajectories, and trains a physics-based tracking policy in simulation. The authors show that a simulated humanoid can then execute lie-and-stand, boxing, and pick-and-place with joint-position mean absolute error of 0.04 to 0.07 meters. They argue this makes humanoid skill acquisition scalable and diverse, cutting the cost and calibration burden of real demonstrations. A sympathetic reader would see this as a step toward training robots from language alone.

Core claim

The paper demonstrates that prompt-conditioned video generation, followed by 3D human pose estimation and motion retargeting, yields reference trajectories that a physics-based imitation-reward RL policy can track successfully in simulation. Evaluated on a humanoid robot across four simulation scenarios, the learned policies complete whole-body transitions and manipulation tasks, preserve multiple natural variations of the same motion, and maintain tracking accuracy under a payload. The authors conclude that this pipeline enables scalable and diverse motion acquisition for complex tasks without relying on motion capture systems or direct human intervention.

What carries the argument

The central mechanism is a three-stage pipeline. First, a text-to-video generation model, guided by a hand-crafted instruction prompt (full-body visibility, static camera, physically plausible motion), produces multiple videos per task. Second, a 3D body model is fit to each video and a retargeting step with inverse kinematics converts the human skeleton into robot-feasible joint trajectories, mitigating foot sliding and ground penetration. Third, a motion-stitching module aligns root poses and interpolates joint angles to chain shorter clips into continuous reference motions. A physics-based RL policy, trained with an imitation reward that balances joint, velocity, end-effector, and center-

Load-bearing premise

The framework relies on the generated videos being physically plausible enough that pose estimation and retargeting produce reference trajectories whose kinematic and dynamic errors are small enough for the imitation reward to drive successful tracking; if the video model produces temporally inconsistent or physically impossible motion, the robot learns biased behavior.

What would settle it

Generate a video for a task the paper flags as challenging, such as a backflip, and run it through the motion-extraction and retargeting pipeline. Measure foot-sliding distance, ground penetration, and center-of-mass acceleration in the resulting reference trajectory; if these exceed the biomechanical plausibility thresholds used for everyday tasks, the claim that generative videos yield robot-feasible motion for dynamic tasks is refuted.

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

If this is right

  • Training data for humanoid skills can be scaled by editing text prompts rather than recording motion capture sessions.
  • A single task description can yield multiple distinct motion styles, giving policies behavioral diversity that is hard to obtain from a few recorded demonstrations.
  • Motion stitching lets one policy learn compound behaviors from independently generated clips, extending the range of tasks without new real data.
  • If the pipeline transfers to physical robots, it could lower the cost of humanoid skill acquisition dramatically, making it accessible to labs without motion-capture facilities.
  • The method's reliance on generative video quality means that tasks with physically demanding or high-speed motions will be the first bottleneck.

Where Pith is reading between the lines

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

  • Editorial inference: the same prompt-to-video-to-motion chain could be applied to other embodied agents (quadrupeds, manipulators) and to generate rare or dangerous motions, like fall recovery, on demand without filming a human.
  • Editorial inference: the reported diversity from multiple generated videos may also act as a regularizer, but it raises a practical question the paper leaves open: how to automatically filter or weight generations that are physically inconsistent.
  • Editorial inference: a testable extension would measure how many generated videos per task are needed to reach a given tracking accuracy, quantifying the trade-off between diversity and reference quality.
  • Editorial inference: if the framework extends to sim-to-real, the physical-consistency errors in generated motion (especially foot sliding and center-of-mass artifacts) would need to be corrected before real-world deployment; the paper identifies this as future work.
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

4 major / 5 minor

Summary. The paper proposes a fully synthetic pipeline for acquiring humanoid control policies: text prompts are fed to Google Veo 3 to generate videos of human motion; SMPL-X pose estimation and GMR retargeting convert these videos into robot reference trajectories; a motion-stitching module concatenates multiple segments; and a DeepMimic-style RL reward, trained with PPO in Isaac Lab, makes a simulated Unitree G1 track the references. Experiments cover lie-and-stand, boxing, and pick-and-place, with quantitative joint-tracking MAE in the range 0.04–0.07 and qualitative torque and end-effector plots. The central claim is that this removes the need for MoCap or real-world demonstrations while enabling scalable, diverse motion acquisition.

Significance. If the central claim holds, the paper points to a scalable alternative to motion capture for humanoid skill acquisition, and it explicitly targets diversity across task variations, which is a recognized weakness of single-demonstration imitation methods. The framework is concrete and reproducible in structure: the RL formulation is standard, the choice of publicly available components (Veo, SMPL-X, GMR, Isaac Lab) is clear, and the paper reports quantitative tracking errors and torque profiles rather than only video stills. The main value is the demonstration that a state-of-the-art video generator can, in at least some tasks, produce references that a physics-based humanoid policy can follow without any robot-specific MoCap. However, the strength of this demonstration is currently limited by missing feasibility checks, missing baselines/ablations, and an overbroad claim about 'no real world data.'

major comments (4)
  1. [III-A, Eq. (5)–(9), V-B, VI] The load-bearing premise—that retargeted synthetic-video references are dynamically feasible for the Unitree G1—is not directly tested. Section III-A states that foot sliding and ground penetration are 'mitigated' by a final IK optimization, but no foot-contact, ground-penetration, center-of-mass, or joint-torque feasibility metrics are reported for the reference trajectories. Section V-B concedes that high-speed segments contain temporal inconsistencies, and Section VI lists 'enhancing physical consistency in generated motions' as future work. Since Eq. (5)–(9) reward tracking of q_ref, infeasible references would bias the policy toward unnatural high-torque compromises. Please report quantitative feasibility metrics for the retargeted trajectories (e.g., foot-contact phase, penetration depth, COM acceleration relative to support polygon) and, if necessary, filter or correct infeasible
  2. [V-C, III-B] The experimental evaluation lacks baselines and ablations that would support the framework's central contribution. There is no comparison against MoCap-based DeepMimic/AMP, another synthetic-data pipeline, or even a simple kinematic-feasibility-filter baseline. Likewise, the motion-stitching module (III-B) is not ablated, and the effect of the number of generated variations per task on policy diversity or success is not measured. Without these comparisons, the observed tracking success cannot be attributed specifically to the generative-video pipeline; it may be an artifact of the RL tracker being robust to almost any smooth reference. Please add at least one MoCap-based baseline and ablation studies for stitching and variation count.
  3. [V-C, Fig. 5, Fig. 6] The quantitative evaluation is too thin to substantiate 'successful execution' and 'strong adaptability.' Only joint-position MAE and selected qualitative plots are given; there are no task success rates, episode completion statistics, failure cases, trials, variance bars, or disturbance-response metrics. The text reports MAE of 'joint positions' in meters (0.04–0.07 m), while Fig. 6's vertical axis label reads 'MAE [rad]', and the figure caption refers to 'joint positions' for what appear to be joint angles; this inconsistency must be resolved. Please define task success operationally and report per-task success rates, distributions over seeds, and tracking errors with confidence intervals.
  4. [Title, Abstract, I, VI] The phrase 'without real world data' in the title and abstract is not supported by the method as described. The video generator (Veo 3), the SMPL-X pose estimator, and the GMR retargeting model are all trained on large collections of real human images/videos. What the paper actually demonstrates is avoiding new robot-specific or task-specific MoCap collection. This distinction is material because the 'without real world data' claim is a headline contribution. Please revise the wording to 'without robot motion-capture data' or 'without new real-world demonstrations' and ensure the abstract/conclusion match this qualified claim.
minor comments (5)
  1. [Fig. 4] The caption says 'four experiments' but only lie-and-stand, boxing, and pick-and-place are named. Either name the fourth experiment or correct the count.
  2. [Fig. 1 vs. V-B] Figure 1 refers to 'Veo 3.1', while Section V-B says 'Google Veo 3 API'. Please unify the model version and explain whether the difference is intentional.
  3. [IV-B, Eq. (13)] The overall reward is written as r_t = r_I_t − r_limit − r_smooth − r_contact. Since the components r_limit, r_smooth, r_contact are described as penalties (positive values), the sign convention is internally consistent, but it would help to state this explicitly and to note whether these terms are bounded or normalized.
  4. [III-B] The motion-stitching description is high-level: what is the duration of the transition buffer, how are interpolation weights chosen, and are joint velocities and accelerations considered? A short formal definition of the stitching operation would improve reproducibility.
  5. [V-A] Minor: 'RTX 4000 Ada GPU' should be more precise (e.g., RTX 4000 Ada Generation), and the PPO hyperparameters are not listed. Provide the learning rate, discount factor, GAE lambda, and training iterations for reproducibility.

Circularity Check

0 steps flagged

No significant circularity: the pipeline is an open-loop composition of external generative, retargeting, and RL components.

full rationale

The derivation chain is open-loop: text prompt → Veo 3 → SMPL-X pose estimation → GMR retargeting → motion stitching → DeepMimic-style RL tracking. Each stage consumes an externally produced signal (video, 3D pose, retargeted reference) rather than a quantity defined by the downstream claim. The reward weights in Eq. (13) are hand-chosen constants, not fitted to the target behaviors, and no equation in Section IV reuses the task-success metric as an input. The reference motion q_ref is not derived from the policy, the reward, or the evaluation score, so no prediction reduces by construction to a fitted value. The only self-citations (refs [2], [3], [10]) appear in the introduction as background and are not load-bearing for the central claim. The evaluation reports tracking error against the same reference trajectories that supervise the policy, which measures fit rather than independent task success; the paper also concedes in the conclusion that 'enhancing physical consistency in generated motions' is future work. These are validation gaps or correctness risks, not circularity: the reference is still an external input, and the claim that the robot can track it is an empirical result, not a tautology. Therefore no circular step is exhibited, and the appropriate score is 0.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

The central pipeline rests on the accuracy of an unshipped chain of proprietary tools (Veo 3, SMPL-X estimator, GMR) plus a set of hand-set reward weights. The reward weights and instruction prompt are the main tunable knobs that are not justified by data.

free parameters (5)
  • Imitation reward weights (w_p, w_v, w_e, w_c) = 0.65, 0.1, 0.15, 0.1
    Hand-chosen weights for Eq. (5); no tuning procedure or sensitivity analysis reported.
  • Exponential tracking scalars (alpha_p, alpha_v, alpha_e, alpha_c) = 2.0, 0.1, 40.0, 10.0
    Hand-set scalings in Eqs. (6-9), controlling how sharply reward decays with tracking error.
  • Regularization weights (w_limit, w_smooth, w_contact) = 1.0, 0.1, 0.1
    Penalty weights in Eqs. (10-12), chosen without reported tuning or ablations.
  • Payload mass in pick-and-place evaluation = 0.5 kg
    Arbitrary test load used in Fig. 5; not swept or motivated.
  • Domain-randomization ranges = not reported
    Section IV-B lists randomized factors (friction, restitution, joint offsets, COM, velocity perturbations) but gives no numerical ranges, so the robustness claim is untestable.
axioms (6)
  • domain assumption Veo 3 generates videos satisfying the instruction prompt's constraints (static camera, full body, physically plausible motion).
    Central to Section III-A; without it, SMPL-X estimation and retargeting are meaningless.
  • domain assumption SMPL-X reconstruction from generated videos is accurate enough for downstream retargeting.
    Section III-A: vision-transformer estimation with bounding-box cues and temporal module; no quantitative validation of reconstruction error on the generated videos is given.
  • domain assumption GMR retargeting plus IK optimization yields dynamically feasible Unitree G1 trajectories that preserve human motion style.
    Section III-A; no evaluation of retargeting error or feasibility beyond downstream policy success.
  • domain assumption Isaac Lab's Unitree G1 dynamics faithfully represent the real robot, so sim-learned policies would transfer.
    Section V-A; paper does not perform sim-to-real transfer, so the framework's applicability to real robots is unestablished.
  • domain assumption The DeepMimic-style reward (Eq. 13) is a sufficient objective for natural task performance.
    Section IV-B; the reward only measures tracking to a reference, not actual task completion, so 'success' in tasks is implicitly defined by tracking.
  • ad hoc to paper The hand-crafted instruction prompt generalizes across all task descriptions used.
    Section III-A and Fig. 2; prompt sensitivity is never explored.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Diverse Humanoid Tasks via Synthetic Video Scenarios without Real World Data." pith.science (2026). https://pith.science/paper/JIPLXROS

@misc{pith2026260721648,
  author       = {Pith},
  title        = {Pith review of: Learning Diverse Humanoid Tasks via Synthetic Video Scenarios without Real World Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JIPLXROS}},
  note         = {Machine review of arXiv:2607.21648}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

The human-like morphology of humanoid robots grants them exceptional potential for agile and versatile motor capabilities, but it also introduces significant challenges in acquiring complex skills. Traditional Learning-from-Demonstrations methods are often constrained by the high cost of collecting real-world data, the difficulty of capturing motion-specific behaviors, and the limited diversity of demonstrations across individuals. Moreover, even for the same task, humans may execute the motion in multiple distinct ways. In this paper, we propose a new framework that leverages the power of Generative AI to convert textual prompts into realistic and diverse sequences of human body movements, enabling the robot to observe multiple variations of how a single task can be performed. These synthetic demonstrations are then used as a training resource, allowing the robot to learn a broad range of task-execution styles without requiring direct human intervention. We evaluate the proposed method across four simulation scenarios. Experimental results show that the robot not only completes the tasks successfully but also demonstrates strong adaptability to complex variations in motion.

Figures

Figures reproduced from arXiv: 2607.21648 by Cong-Thanh Vu, Yen-Chen Liu, Yun-Hao Tsai.

Figure 1
Figure 1. Figure 1: Overview of the proposed framework. Multi-modal prompts guide the Veo 3.1 model to generate human motion videos, which are converted into [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: Human skeletal motion is extracted from Generative AI video data and [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative results from four experiments evaluating policies trained using generative AI with a single prompt. All motions are executed on the humanoid [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Evaluation of the learned policy for the box pick-and-place task, showing upper- and lower-body joint positions, vertical trajectories of the left wrist [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: MAE of joint positions for both upper and lower body across the [PITH_FULL_IMAGE:figures/full_fig_p006_6.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

23 extracted references · 6 linked inside Pith

  1. [1]

    Development of a multiple-sensor navigation system for autonomous guided vehicle localization,

    A.-T. Nguyen, V .-T. Nguyen, X.-T. Nguyen, and C.-T. Vu, “Development of a multiple-sensor navigation system for autonomous guided vehicle localization,” inIntelligent Systems and Networks(D.-T. Tran, G. Jeon, T. D. L. Nguyen, J. Lu, and T.-D. Xuan, eds.), (Singapore), pp. 402–410, Springer Singapore, 2021

    2021

  2. [2]

    Autonomous adjustment of tracking position in dynamic environments for human-following robots using deep reinforce- ment learning,

    C.-T. Vu and Y .-C. Liu, “Autonomous adjustment of tracking position in dynamic environments for human-following robots using deep reinforce- ment learning,” in2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 16863–16869, 2025

    2025

  3. [3]

    Autonomous navigation for human-following robots based on optimized position tracking,

    C.-T. Vu, H.-H. Huang, and Y .-C. Liu, “Autonomous navigation for human-following robots based on optimized position tracking,” in2025 10th International Conference on Control and Robotics Engineering (ICCRE), pp. 23–27, 2025

    2025

  4. [4]

    Opportunities challenges and roadmap for humanoid robots in construction,

    T. Uthai, H. You, M. Wang, K. Smith, E. Spackman, Z. Ryan, S. Li, and J. Du, “Opportunities challenges and roadmap for humanoid robots in construction,”Scientific Reports, 2025

    2025

  5. [5]

    Humanoid locomotion and manipulation: Current progress and challenges in control, planning, and learning,

    Z. Gu, J. Li, W. Shen, W. Yu, Z. Xie, S. McCrory, X. Cheng, A. Shamsah, R. Griffin, C. K. Liu,et al., “Humanoid locomotion and manipulation: Current progress and challenges in control, planning, and learning,”arXiv preprint arXiv:2501.02116, 2025

    Pith/arXiv arXiv 2025

  6. [6]

    Passive dynamic walking,

    T. McGeer, “Passive dynamic walking,”The international journal of robotics research, vol. 9, no. 2, pp. 62–82, 1990

    1990

  7. [7]

    Passive dynamic walking with symmetric fixed flat feet,

    J. Kim, C.-H. Choi, and M. W. Spong, “Passive dynamic walking with symmetric fixed flat feet,” in2007 IEEE International Conference on Control and Automation, pp. 24–30, 2007

    2007

  8. [8]

    A necessary condition for passive dynamic walking,

    Y . Iwatani and T. Kinugasa, “A necessary condition for passive dynamic walking,” in2022 American Control Conference (ACC), pp. 1885–1890, 2022

    2022

  9. [9]

    Towards active actuated natural walking humanoid robot legs,

    R. C. Luo, C. H. Chen, Y . H. Pu, and J. R. Chang, “Towards active actuated natural walking humanoid robot legs,” in2011 IEEE/ASME International Conference on Advanced Intelligent Mechatronics (AIM), pp. 886–891, 2011

    2011

  10. [10]

    Learning virtual passive dynamic walking using the kneed walker model for guiding policies,

    C.-T. Vu, C.-C. Lai, and Y .-C. Liu, “Learning virtual passive dynamic walking using the kneed walker model for guiding policies,” in2025 IEEE-RAS 24th International Conference on Humanoid Robots (Hu- manoids), pp. 1094–1100, 2025

    2025

  11. [11]

    Beyondmimic: From motion tracking to versatile humanoid control via guided diffusion,

    Q. Liao, T. E. Truong, X. Huang, Y . Gao, G. Tevet, K. Sreenath, and C. K. Liu, “Beyondmimic: From motion tracking to versatile humanoid control via guided diffusion,”arXiv preprint arXiv:2508.08241, 2025

    Pith/arXiv arXiv 2025

  12. [12]

    Advancements in humanoid robots: A comprehensive review and future prospects,

    Y . Tong, H. Liu, and Z. Zhang, “Advancements in humanoid robots: A comprehensive review and future prospects,”IEEE/CAA Journal of Automatica Sinica, vol. 11, no. 2, pp. 301–328, 2024

    2024

  13. [13]

    Ex- pressive whole-body control for humanoid robots,

    X. Cheng, Y . Ji, J. Chen, R. Yang, G. Yang, and X. Wang, “Ex- pressive whole-body control for humanoid robots,”arXiv preprint arXiv:2402.16796, 2024

    Pith/arXiv arXiv 2024

  14. [14]

    Deepmimic: example-guided deep reinforcement learning of physics-based character skills,

    X. B. Peng, P. Abbeel, S. Levine, and M. van de Panne, “Deepmimic: example-guided deep reinforcement learning of physics-based character skills,”ACM Trans. Graph., vol. 37, July 2018

    2018

  15. [15]

    Amp: adversarial motion priors for stylized physics-based character control,

    X. B. Peng, Z. Ma, P. Abbeel, S. Levine, and A. Kanazawa, “Amp: adversarial motion priors for stylized physics-based character control,” ACM Trans. Graph., vol. 40, July 2021

    2021

  16. [16]

    I-ctrl: Imitation to control humanoid robots through bounded residual reinforcement learning,

    Y . Yan, E. V . Mascaro, T. Egle, and D. Lee, “I-ctrl: Imitation to control humanoid robots through bounded residual reinforcement learning,”IEEE Robotics & Automation Magazine, vol. 32, no. 1, pp. 59–67, 2025

    2025

  17. [17]

    Dfm: Deep fourier mimic for ex- pressive dance motion learning,

    R. Watanabe, C. Li, and M. Hutter, “Dfm: Deep fourier mimic for ex- pressive dance motion learning,” in2025 IEEE International Conference on Robotics and Automation (ICRA), pp. 9644–9650, 2025

    2025

  18. [18]

    Learning aggressive animal locomotion skills for quadrupedal robots solely from monocular videos,

    L. Zhao, Z. Luo, Y . Han, J. Zhang, Y . Chen, Y . Liu, and P. Lu, “Learning aggressive animal locomotion skills for quadrupedal robots solely from monocular videos,”npj Robotics, vol. 3, no. 1, p. 32, 2025

    2025

  19. [19]

    Hunyuanvideo: A systematic framework for large video generative models,

    W. Kong, Q. Tian, Z. Zhang, R. Min, Z. Dai, J. Zhou, J. Xiong, X. Li, B. Wu, J. Zhang,et al., “Hunyuanvideo: A systematic framework for large video generative models,”arXiv preprint arXiv:2412.03603, 2024

    Pith/arXiv arXiv 2024

  20. [20]

    Wan: Open and advanced large-scale video generative models,

    T. Wan, A. Wang, B. Ai, B. Wen, C. Mao, C.-W. Xie, D. Chen, F. Yu, H. Zhao, J. Yang,et al., “Wan: Open and advanced large-scale video generative models,”arXiv preprint arXiv:2503.20314, 2025

    Pith/arXiv arXiv 2025

  21. [21]

    Expressive body capture: 3d hands, face, and body from a single image,

    G. Pavlakos, V . Choutas, N. Ghorbani, T. Bolkart, A. A. Osman, D. Tzionas, and M. J. Black, “Expressive body capture: 3d hands, face, and body from a single image,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 10975–10985, 2019

    2019

  22. [22]

    Retargeting matters: General motion retargeting for humanoid motion tracking,

    J. P. Araujo, Y . Ze, P. Xu, J. Wu, and C. K. Liu, “Retargeting matters: General motion retargeting for humanoid motion tracking,”arXiv preprint arXiv:2510.02252, 2025

    arXiv 2025

  23. [23]

    Prox- imal policy optimization algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Prox- imal policy optimization algorithms,”arXiv preprint arXiv:1707.06347, 2017

    Pith/arXiv arXiv 2017

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.