Pith. sign in

REVIEW 3 major objections 6 minor 24 references

CORB-Planner claims that a single RL policy, trained in about ten minutes on a simplified obstacle course using only safe-flight-corridor observations, can plan real-time B-spline trajectories for quadrotors, hexarotors, LiDAR, and vision p

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 →

CORB-Planner uses safe flight corridors as low-dimensional observations for an RL policy that generates B-spline control points, enabling real-time cross-platform UAV planning after about ten minutes of training.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection A credible RL-based B-spline planner with real cross-platform flights, but the sim-to-real claim leans on an unvalidated perfect-tracking assumption and the abstract oversells the lightweight board's speed. the 3 major comments →

arxiv 2509.11240 v1 pith:JLBFWJKF submitted 2025-09-14 cs.RO cs.SYeess.SY

CORB-Planner: Corridor as Observations for RL Planning in High-Speed Flight

classification cs.RO cs.SYeess.SY
keywords reinforcement learningsafe flight corridorB-spline trajectory planningsim-to-real transferUAV autonomyhigh-speed flightonboard planningSDCQ
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

CORB-Planner sets out to solve the sim-to-real problem for RL-based UAV trajectory planning by making the observation space platform-independent: instead of raw LiDAR or depth images, the policy sees a safe flight corridor built around an A* reference path, plus recent B-spline control points and a time index. The policy outputs one acceleration control point per step, which becomes the next B-spline control point, so the planner produces a smooth, dynamically feasible trajectory that the UAV's low-level controller tracks. Training is done in a deliberately simplified simulation without physics, using an easy-to-hard wall course, an exploitation-decoupled sampling scheme, and a soft decomposed-critic Q (SDCQ) value algorithm; the paper reports convergence in about ten minutes. If correct, a single policy can transfer across airframes and sensor suites and run on a tiny onboard computer, reaching 8.2 m/s in cluttered real-world flights. The paper's central claim is that low-dimensional SFC observations, not sensor-level details, are the right substrate for reusable RL planners in high-speed flight.

Core claim

The central claim: a safe flight corridor (SFC) can be a compact, platform-independent observation for an RL policy that plans B-spline trajectories. Around an A* reference polyline, SFC is a sequence of rectangular sub-corridors, each described by eight numbers (endpoints, left/right obstacle distances, z-range), so under 100 features encode the obstacle layout. The policy takes the nine nearest sub-corridors plus three recent control points and time (66 dimensions) and emits an acceleration control point in a cylindrical, uniformly constrained space, which becomes the next B-spline control point. Training uses the SDCQ value algorithm, a difficulty-progressive course, and multiple explorat

What carries the argument

The central object is the safe flight corridor as a low-dimensional observation. SFC is built by pruning an A* path into a minimal polyline, then attaching left and right rectangular sub-corridors whose widths record the distance to the nearest obstacle on each side within a chosen z-band. Each sub-corridor is described by eight numbers, so a local window of nine sub-corridors plus recent control points yields a 66-dimensional state. The companion mechanism is B-spline continuation: the RL action is a bounded acceleration control point, mapped through a cylindrical normalization to enforce uniform acceleration limits, then integrated to produce the next position control point; the third-orde

Load-bearing premise

The load-bearing assumption is that any UAV with a capable low-level controller tracks the planned B-splines almost perfectly, because the training simulation omits physics entirely and says 'the tracking of B-spline trajectories is assumed to be perfectly precise' (Section IV-B); if tracking lag or platform-specific dynamics violate this, the ten-minute sim-to-real transfer and the 8.2 m/s results do not follow.

What would settle it

Take the vmax=10 policy and fly it in the same cluttered environment while deliberately adding 100 ms of tracking latency or lowering the controller gains until the platform noticeably lags the B-spline. If success rate collapses while the planning inputs are unchanged, the perfect-tracking premise is the load-bearing link. A cleaner computational test: run a single realistic quadrotor simulator with imperfect tracking but the identical SFC observations and measure whether the policy's success rate falls below the no-physics training result.

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

If this is right

  • Deploying CORB-Planner on a new vehicle does not require retraining: the same agents trained at maximum velocities of 4, 5, 7, 10, and 15 m/s ran on quadrotors, a hexarotor, LiDAR, and vision platforms in the reported experiments.
  • Because the policy outputs only B-spline control points, onboard planning cost is small enough to run at 25 Hz on a 275 g quadrotor with a small ARM board, leaving most CPU capacity to odometry.
  • Training is fast and cheap, about ten minutes on a desktop GPU without physics simulation, which could make RL planning practical for rapid iteration in field robotics.
  • On dense-wall courses, the aggressive variant of CORB-Planner is faster but less reliable than the conservative variant, so choosing reward weights trades speed against success rate.
  • Slowly moving obstacles can be handled by updating the occupancy grid and commanding a hover when no A* path exists, as shown in the moving-forest test.

Where Pith is reading between the lines

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

  • If the SFC representation is the true source of generality, then an optimization-based planner given the same 66-dimensional corridor observation should also transfer across platforms; testing that would separate the contribution of the representation from the contribution of RL.
  • The perfect-tracking training assumption implies a sharp test: degrade the low-level tracker by adding latency or lowering gains and measure how success rate drops; the 8.2 m/s figure should be read as specific to the tested low-level controller, not a platform-independent ceiling.
  • Because the action space enforces uniform horizontal acceleration, the policy is likely to generalize to vehicles with different mass and rotor layouts, but vehicles with asymmetric actuation limits would require a different action transform; extending the cylindrical mapping to those cases is a natural next step.
  • The moving-obstacle experiments were limited to slow pedestrians; extending CORB-Planner to faster dynamic obstacles would require the SFC to encode velocity information, which the current 66-dimensional state does not include.
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

3 major / 6 minor

Summary. The paper presents CORB-Planner, a real-time RL-based B-spline trajectory planner for UAVs. The observation is a compact safe flight corridor (SFC) derived from an A* reference polyline, together with the current B-spline control points and time. An RL policy outputs acceleration control points that are transformed into B-spline control points; a reward function penalizes SFC violations, rewards progress along the SFC, and applies a jerk discount. Training uses the authors' SDCQ algorithm in a physics-free easy-to-hard curriculum with exploitation-decoupled multi-thread sampling, claiming convergence in about ten minutes. Experiments compare against EGO-planner in simulation, ablate algorithm/training choices, and report real-world flights on four UAV platforms, including an 8.2 m/s forest flight and a 275 g quadrotor with a Cortex-A53 board.

Significance. If the central claims hold, the paper demonstrates a practical route to cross-platform, low-compute RL trajectory planning: a single simulation-trained policy transfers to quadrotors, hexarotors, LiDAR, and vision sensing, and runs at 25 Hz on a lightweight board. The SFC-as-observation idea is a clean and credible mechanism for reducing overfitting to platform-specific sensing and dynamics. The paper provides external baselines (EGO-planner, DPMPC-planner), ablations against PPO/SAC, and physical experiments on multiple platforms, which are significant assets. However, the absence of code/data and the unvalidated perfect-tracking assumption weaken the reproducibility and quantitative support for the cross-platform claim.

major comments (3)
  1. [Section IV-B, with Section V-D/E] The training environment 'does not incorporate physical simulations' and 'the tracking of B-spline trajectories is assumed to be perfectly precise'; only random noise is added to the planning process. The real-world transfer claims (Section V-D and the 8.2 m/s forest flight in Section V-E) rest on the onboard PID+PX4 controller tracking the generated B-spline closely enough that the vehicle remains inside the SFC. No tracking-error measurement, SFC clearance margin, or sensitivity analysis is reported. If the tracking error is comparable to the corridor's free space, the policy decisions can become infeasible. Please report measured position tracking error versus SFC widths/clearances for each real-world platform and speed, and/or add a tracking-error model or dynamics noise in training and re-evaluate. This is load-bearing for the central cross-platform and high-speed claims.
  2. [Algorithm 1 in Section III-A] As printed, line 7 'Add u_k to u*' is inside the for loop, so the final waypoint u_k is appended at every iteration, producing a corrupted polyline. The polyline directly determines the SFC used in the observation and reward, so this is not merely cosmetic. Please correct the pseudocode by moving line 7 after the loop and verify the loop range, which currently starts at i=2 and appears to skip u_1.
  3. [Section IV-A, Eq. (18)] The discrete action index is defined as α_d^t ∈ [1,M]^3 and the mapping is α_t = (2α_d+1)/M−1. For α_d=M and M=60, this gives α_t ≈ 1.0167, outside the claimed cubic action space [−1,1]^3; for α_d=1 it gives −0.95. If implemented literally, the action bounds are inconsistent with Section III-B. Please clarify the index convention (e.g., α_d ∈ [0,M−1] with mapping (2α_d+1)/M−1, or use 2(α_d−1)/(M−1)−1 to exactly cover [−1,1]).
minor comments (6)
  1. [Section III-C] The text introduces coefficients 'k_p, k_v, and k_s', but Eq. (9) uses k_f. Please unify the notation.
  2. [Section III-A, Eq. (4)] The orthogonal vector n_i contains subscript typos: 'u^*_{i,y-1}' and 'u_{i,y}' should likely be 'u^*_{i-1,y}' and 'u^*_{i,y}'. Please fix.
  3. [Section V-A] The vision-based quadrotor is described with a 'RealSense D430' here, while the Introduction and Section V-D say 'RealSense D435i'. Please make the sensor name consistent.
  4. [Section V-A] The text refers to 'Fig. I' but should refer to 'Fig. 1'.
  5. [Reference [14]] The cited title is 'Soft Decomposed Policy-Critic', but the paper describes 'Soft Decomposed-Critic Q (SDCQ)'. Please align the reference title with the algorithm name.
  6. [General] No code or data release is mentioned. Given the ten-minute training claim and the cross-platform transfer result, a reproducibility statement or supplementary material (e.g., trained policy parameters, training configuration, and a sample real-world dataset) would substantially strengthen the paper.

Circularity Check

0 steps flagged

No significant circularity: the central claims are experimentally validated against external baselines and physical platforms; self-citations are not load-bearing.

full rationale

The paper's core derivation chain is not circular. The RL policy is trained to output B-spline control points using an SFC-based observation and a reward that encourages SFC-following, but the paper's claims concern measured navigation performance: success rates, episode times, obstacle avoidance, and real-world flights. These outcomes are empirically compared against external baselines (EGO-planner, DPMPC-planner, PPO, SAC), not derived from the training reward or the observation representation. The SDCQ algorithm is cited to the authors' prior work [14], but the paper states the loss functions (Eqs. 18-21), the discretization scheme, and provides an ablation against PPO and SAC, so the 'ten-minute training' claim rests on in-paper experiments rather than on the self-citation alone. Reference [22] is from overlapping authors but is only mentioned in related work and is not load-bearing. Section IV-B explicitly acknowledges the simplifying assumption of perfectly precise B-spline tracking and the absence of physics simulation; this is a stated limitation and a robustness concern for sim-to-real transfer, but it is not circular reasoning, because the real-world deployment results provide an independent, falsifiable check. No fitted parameter is renamed as a prediction: the reward weights are hand-tuned and reported, and the policy's outputs are measured, not algebraically forced by the inputs. No uniqueness theorem from prior work is invoked to exclude alternatives. The only self-citations are non-load-bearing, hence the low score.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

No new physical entities are introduced. The free parameters are engineering choices: reward weights, speed-scaling formulas, SFC horizon, discretization level, and collision sampling. The load-bearing assumptions are the sufficiency of the SFC representation, the perfect-tracking premise in training, and the borrowed SDCQ algorithm. The central contribution is an empirical integration, so the ledger consists mostly of domain assumptions rather than invented physics.

free parameters (5)
  • reward weights k_p, k_f, k_s = main: -30, 5, 50; fast variant: -30, 8; safe variant: -50, 3
    Manually chosen to balance safety, speed, and success; no sensitivity analysis or tuning procedure is given.
  • acceleration and jerk scaling relative to v_max = a_max = 2*v_max; j_max = 50 + 10*v_max
    Fixed formulas chosen by hand per speed class; they define the action space and the jerk termination threshold.
  • SFC horizon N in the observation = N = 9 sub-corridors, giving a 66-dimensional state
    Set 'in practice'; no ablation or justification that 9 corridors preserve decision-relevant information.
  • action discretization level M in SDCQ = M = 60
    Selected after ablation as the accuracy/efficiency trade-off; no formal criterion is reported.
  • collision sampling count in the reward = 10 samples per knot interval
    Chosen for the SFC-violation check in Eq. (12); under-sampling could miss violations between samples.
axioms (5)
  • standard math Uniform third-order B-spline basis recursion, local support, and convex hull properties hold as stated in Eq. (1).
    Used in Section II-B to define the trajectory and to derive velocity, acceleration, and jerk control points.
  • domain assumption A cubic SFC built around a shortened A* polyline is a sufficient low-dimensional summary of the obstacle environment for policy decisions.
    Section III-A, Eqs. (5)-(6): the SFC compresses the grid map to 66 numbers; no information-loss analysis is provided.
  • domain assumption The trained policy transfers to real hardware because low-level controllers track the B-spline precisely; the simulator contains no physics and adds noise only to the planning process.
    Section IV-B explicitly assumes perfect tracking; random noise simulates planning errors, not dynamics or tracking lag.
  • domain assumption SDCQ as specified in [14] converges to a useful policy in about ten minutes under the authors' curriculum and sampling scheme.
    The present paper provides no proof, no full algorithm details, and no released code; it relies on the authors' prior result [14].
  • domain assumption Localization (Fast-LIO2, Point-LIO, VINS-Fusion) and grid-map resolution (0.15 m) are accurate enough for safe planning on every tested platform.
    Section V-A/V-D attribute failures to VIO quality, implying performance depends on localization accuracy, which is treated as an external input.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of CORB-Planner: Corridor as Observations for RL Planning in High-Speed Flight." pith.science (2026). https://pith.science/paper/JLBFWJKF

@misc{pith2026250911240,
  author       = {Pith},
  title        = {Pith review of: CORB-Planner: Corridor as Observations for RL Planning in High-Speed Flight},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JLBFWJKF}},
  note         = {Machine review of arXiv:2509.11240}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Reinforcement learning (RL) has shown promise in a large number of robotic control tasks. Nevertheless, its deployment on unmanned aerial vehicles (UAVs) remains challenging, mainly because of reliance on accurate dynamic models and platform-specific sensing, which hinders cross-platform transfer. This paper presents the CORB-Planner (Corridor-as-Observations for RL B-spline planner), a real-time, RL-based trajectory planning framework for high-speed autonomous UAV flight across heterogeneous platforms. The key idea is to combine B-spline trajectory generation with the RL policy producing successive control points with a compact safe flight corridor (SFC) representation obtained via heuristic search. The SFC abstracts obstacle information in a low-dimensional form, mitigating overfitting to platform-specific details and reducing sensitivity to model inaccuracies. To narrow the sim-to-real gap, we adopt an easy-to-hard progressive training pipeline in simulation. A value-based soft decomposed-critic Q (SDCQ) algorithm is used to learn effective policies within approximately ten minutes of training. Benchmarks in simulation and real-world tests demonstrate real-time planning on lightweight onboard hardware and support maximum flight speeds up to 8.2m/s in dense, cluttered environments without external positioning. Compatibility with various UAV configurations (quadrotors, hexarotors) and modest onboard compute underlines the generality and robustness of CORB-Planner for practical deployment.

Figures

Figures reproduced from arXiv: 2509.11240 by Bin Gao, Gang Wang, Jian Sun, Yechen Zhang, Zhuo Li.

Figure 1
Figure 1. Figure 1: System architecture of CORB-Planner. The preprocessing module constructs the SFC and B-spline initialization from onboard odometry and grid [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: An example of A∗ path u, reference polyline u ∗ and the SFC defined in CORB-Planner. Top view of the example is on the right side of the figure. SF Cl i on the left side of the polyline is represented by the orange cubic, and SF Cr i on the right side of the polyline by the blue cubic. Algorithm 1 Get polyline u ∗ from A* path u Require: grid map, u =< u0 → u1 → · · · → uk > 1: Initialize u ∗ =< u0 >, loca… view at source ↗
Figure 3
Figure 3. Figure 3: Training system of CORB-Planner, enables multi-thread sampling and experience replay. With only [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Behaviors of both CORB-Plnnaer and EGO-planner in the random simulation environments with maximum velocity [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Physical experiments of CORB-Planner. Indoor Experiments: (A) High-performance quadrotor with the [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Results of ablation studies: (A) comparison of RL algorithms [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: CORB-Planner on the ultra lightweight quadrotor. (A) physical [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
Figure 7
Figure 7. Figure 7: (A) Physical performance of moving obstacle avoidance. [PITH_FULL_IMAGE:figures/full_fig_p010_7.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

24 extracted references · 6 linked inside Pith

  1. [1]

    Learning agile soccer skills for a bipedal robot with deep reinforcement learning,

    T. Haarnoja, B. Moran, G. Lever, S. H. Huang, D. Tirumala, J. Humplik, M. Wulfmeier, S. Tunyasuvunakool, N. Y . Siegel, R. Hafneret al., “Learning agile soccer skills for a bipedal robot with deep reinforcement learning,”Science Robotics, vol. 9, no. 89, p. eadi8022, 2024

  2. [2]

    Learning quadrupedal locomotion over challenging terrain,

    J. Lee, J. Hwangbo, L. Wellhausen, V . Koltun, and M. Hutter, “Learning quadrupedal locomotion over challenging terrain,”Science Robotics, vol. 5, no. 47, p. eabc5986, 2020

  3. [3]

    Champion-level drone racing using deep reinforcement learning,

    E. Kaufmann, L. Bauersfeld, A. Loquercio, M. M ¨uller, V . Koltun, and D. Scaramuzza, “Champion-level drone racing using deep reinforcement learning,”Nature, vol. 620, no. 7976, pp. 982–987, 2023

  4. [4]

    Back to Newton’s laws: Learning vision-based agile flight via differentiable physics,

    Y . Zhang, Y . Hu, Y . Song, D. Zou, and W. Lin, “Back to Newton’s laws: Learning vision-based agile flight via differentiable physics,”arXiv preprint arXiv:2407.10648, 2024

  5. [5]

    Efficient and robust time-optimal trajectory planning and control for agile quadrotor flight,

    Z. Zhou, G. Wang, J. Sun, J. Wang, and J. Chen, “Efficient and robust time-optimal trajectory planning and control for agile quadrotor flight,” IEEE Robot. Autom. Lett., vol. 8, no. 12, pp. 7913–7920, 2023

  6. [6]

    Learning high-speed flight in the wild,

    A. Loquercio, E. Kaufmann, R. Ranftl, M. M ¨uller, V . Koltun, and D. Scaramuzza, “Learning high-speed flight in the wild,”Science Robotics, vol. 6, no. 59, p. eabg5810, 2021. 11

  7. [7]

    Fixed-time disturbance observer-based MPC robust trajectory tracking control of quadrotor,

    L. Xu, B. Tian, C. Wang, J. Lu, D. Wang, Z. Li, and Q. Zong, “Fixed-time disturbance observer-based MPC robust trajectory tracking control of quadrotor,”IEEE/ASME Trans. Mech., pp. 1–11, 2024, DOI:10.1109/TMECH.2024.3503062

  8. [8]

    Discrete-time adaptive neural tracking control and its experiments for quadrotor unmanned aerial vehicle systems,

    X. Zhang, Y . Wang, G. Zhu, X. Chen, and C.-Y . Su, “Discrete-time adaptive neural tracking control and its experiments for quadrotor unmanned aerial vehicle systems,”IEEE/ASME Trans. Mech., vol. 28, no. 3, pp. 1201–1212, 2023

  9. [9]

    A safety planning and control architecture applied to a quadrotor autopilot,

    W. Zhang, J. Jia, S. Zhou, K. Guo, X. Yu, and Y . Zhang, “A safety planning and control architecture applied to a quadrotor autopilot,”IEEE Robot. Autom. Lett., vol. 8, no. 2, pp. 680–687, 2022

  10. [10]

    FAST- LIVO2: Fast, direct LiDAR–inertial–visual odometry,

    C. Zheng, W. Xu, Z. Zou, T. Hua, C. Yuan, D. He, B. Zhou, Z. Liu, J. Lin, F. Zhu, Y . Ren, R. Wang, F. Meng, and F. Zhang, “FAST- LIVO2: Fast, direct LiDAR–inertial–visual odometry,”IEEE Trans. Robot., vol. 41, pp. 326–346, 2025

  11. [11]

    Point-LIO: Robust high-bandwidth light detection and ranging inertial odometry,

    D. He, W. Xu, N. Chen, F. Kong, C. Yuan, and F. Zhang, “Point-LIO: Robust high-bandwidth light detection and ranging inertial odometry,” Adv. Intell. Syst., vol. 5, no. 7, p. 2200459, 2023

  12. [12]

    A general optimization-based framework for global pose estimation with multiple sensors,

    T. Qin, S. Cao, J. Pan, and S. Shen, “A general optimization-based framework for global pose estimation with multiple sensors,”arXiv preprint arXiv:1901.03642, 2019

  13. [13]

    Planning dynamically feasible trajectories for quadrotors using safe flight corridors in 3D complex environments,

    S. Liu, M. Watterson, K. Mohta, K. Sun, S. Bhattacharya, C. J. Taylor, and V . Kumar, “Planning dynamically feasible trajectories for quadrotors using safe flight corridors in 3D complex environments,”IEEE Robot. Autom. Lett., vol. 2, no. 3, pp. 1688–1695, 2017

  14. [14]

    Soft decomposed policy-critic: Bridging the gap for effective continuous control with discrete RL,

    Y . Zhang, J. Sun, G. Wang, Z. Li, and W. Chen, “Soft decomposed policy-critic: Bridging the gap for effective continuous control with discrete RL,”arXiv:2308.10203, 2023

  15. [15]

    Ego-planner: An ESDF- free gradient-based local planner for quadrotors,

    X. Zhou, Z. Wang, H. Ye, C. Xu, and F. Gao, “Ego-planner: An ESDF- free gradient-based local planner for quadrotors,”IEEE Robot. Autom. Lett., vol. 6, no. 2, pp. 478–485, 2020

  16. [16]

    Proximal policy optimization algorithms,

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

  17. [17]

    Continuous control with deep reinforcement learning,

    T. P. Lillicrap, J. J. Hunt, A. Pritzel, N. Heess, T. Erez, Y . Tassa, D. Silver, and D. Wierstra, “Continuous control with deep reinforcement learning,”arXiv:1509.02971, 2015

  18. [18]

    Addressing function approxi- mation error in actor-critic methods,

    S. Fujimoto, H. Hoof, and D. Meger, “Addressing function approxi- mation error in actor-critic methods,” inProc. Int. Conf. Mach. Learn. PMLR, 2018, pp. 1587–1596

  19. [19]

    Soft actor-critic: Off- policy maximum entropy deep reinforcement learning with a stochastic actor,

    T. Haarnoja, A. Zhou, P. Abbeel, and S. Levine, “Soft actor-critic: Off- policy maximum entropy deep reinforcement learning with a stochastic actor,” inProc. Int. Conf. Mach. Learn.PMLR, 2018, pp. 1861–1870

  20. [20]

    Soft actor-critic algorithms and applications,

    T. Haarnoja, A. Zhou, K. Hartikainen, G. Tucker, S. Ha, J. Tan, V . Ku- mar, H. Zhu, A. Gupta, P. Abbeelet al., “Soft actor-critic algorithms and applications,”arXiv:1812.05905, 2018

  21. [21]

    Mastering diverse control tasks through world models,

    D. Hafner, J. Pasukonis, J. Ba, and T. Lillicrap, “Mastering diverse control tasks through world models,”Nature, pp. 1–7, 2025

  22. [22]

    Storm: Efficient stochastic transformer based world models for reinforcement learning,

    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., vol. 36, 2023, pp. 27 147–27 166

  23. [23]

    Bubble planner: Planning high-speed smooth quadrotor trajectories using receding corridors,

    Y . Ren, F. Zhu, W. Liu, Z. Wang, Y . Lin, F. Gao, and F. Zhang, “Bubble planner: Planning high-speed smooth quadrotor trajectories using receding corridors,” inIEEE/RSJ Int. Conf . Intell. Robots Syst. IEEE, 2022, pp. 6332–6339

  24. [24]

    DPMPC-Planner: A real- time UA V trajectory planning framework for complex static environ- ments with dynamic obstacles,

    Z. Xu, D. Deng, Y . Dong, and K. Shimada, “DPMPC-Planner: A real- time UA V trajectory planning framework for complex static environ- ments with dynamic obstacles,” inIEEE Int. Conf. Robot. Autom.IEEE, 2022, pp. 250–256

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.