pith. sign in

arxiv: 2606.28827 · v1 · pith:VRH2WRGFnew · submitted 2026-06-27 · 💻 cs.RO

LNN-Fly: Continuous-Time UAV Navigation for Robust Obstacle Avoidance under Timing Mismatch

Yulin Huang , Shaojie Chen , Di Feng , Jiahao Wang , Ping Liu , Jianxiao Zou This is my paper

Pith reviewed 2026-06-30 09:53 UTC · model grok-4.3

classification 💻 cs.RO
keywords UAV navigationobstacle avoidancecontinuous-time policytiming mismatchLiDARsim-to-real transferquadrotorrecurrent update
0
0 comments X

The pith

A continuous-time UAV policy with explicit interval conditioning and adaptive forgetting maintains obstacle avoidance under timing mismatches and stale observations.

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

The paper presents LNN-Fly, a LiDAR-based navigation policy for UAVs designed to tolerate irregular control updates and delayed sensor data that commonly break deployed policies. It employs a structured recurrent update inspired by dynamic programming, conditions directly on the elapsed control interval, and uses an adaptive forgetting gate to refresh outdated state information near hazards. Training occurs via differentiable rollouts that inject realistic sensing and timing perturbations, yielding improved robustness in simulation and direct transfer to hardware. A reader would care because these timing effects are frequent failure modes when moving agile navigation policies from simulation into cluttered real environments.

Core claim

LNN-Fly combines a dynamic-programming-inspired structured recurrent update, explicit conditioning on the elapsed control interval Δt, and an input-driven adaptive forgetting gate that refreshes stale latent state near hazards while preserving consistency during sustained maneuvers. Trained with differentiable rollouts that incorporate deployment-relevant sensing and timing perturbations, the policy improves obstacle-avoidance performance under reduced control frequency, sparse observations, and control-period jitter. It transfers zero-shot to a physical quadrotor and achieves 100% success over 20 flights in indoor cross-frequency real-world tests.

What carries the argument

Dynamic-programming-inspired structured recurrent update with explicit Δt conditioning and input-driven adaptive forgetting gate that manages latent state consistency under variable timing.

If this is right

  • Improves obstacle-avoidance performance under reduced control frequency, sparse observations, and control-period jitter in simulation.
  • Transfers zero-shot from a simplified differentiable simulator to a physical quadrotor.
  • Achieves 100% success over 20 flights in indoor cross-frequency real-world tests.
  • Maintains low inference latency with median 0.514 ms on GPU, about 2.5 ms on CPU, and onboard P95 below 30 ms.

Where Pith is reading between the lines

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

  • The training approach of injecting explicit timing perturbations could apply to other continuous-control robots that experience variable update rates.
  • The adaptive forgetting gate may extend usefully to policies using sensors other than LiDAR when input staleness is an issue.
  • Success of zero-shot transfer suggests that modeling deployment mismatches in simulation can lessen dependence on post-training adaptation.
  • Further tests in outdoor or wind-affected settings would reveal whether the timing robustness generalizes beyond controlled indoor flights.

Load-bearing premise

The sensing irregularity and variable-rate control perturbations added during training are representative of the actual timing mismatches and observation staleness that occur on the physical platform.

What would settle it

Real-world flights under deliberately varied control periods and reduced observation rates that produce any collisions would show the robustness does not hold.

Figures

Figures reproduced from arXiv: 2606.28827 by Di Feng, Jiahao Wang, Jianxiao Zou, Ping Liu, Shaojie Chen, Yulin Huang.

Figure 1
Figure 1. Figure 1: Real-world evaluations in diverse environments. Subfigures (a) and (b) show real-flight experiments in an outdoor forest environment, while (c) and (d) show real-flight experiments in a narrow indoor laboratory corridor. The results demonstrate robust navigation performance across highly cluttered natural terrain and confined man-made spaces. collision-free navigation in cluttered environments under real o… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the LNN-Fly framework. The system processes LiDAR sectors [41] and state cues via a lightweight encoder and a Closed-form Continuous-time (CfC) core. An input-driven forgetting mechanism dynamically modulates the temporal scale to balance smoothness and agility, optimized end-to-end via differentiable physics. belief surrogate hk to actions uk while remaining robust to these discrepancies. To t… view at source ↗
Figure 3
Figure 3. Figure 3: Visualization of the adaptive forgetting rate λk . The trajectory color transitions from blue (low λk ) to red (high λk ). During sustained avoidance of a large obstacle, λk remains low to preserve memory and maintain a consistent avoidance direction. After the obstacle is cleared, λk spikes to rapidly refresh the latent state for the subsequent flight segment. the need to trade off memory retention and ra… view at source ↗
Figure 4
Figure 4. Figure 4: Benchmark results with previous systems under different maximum speeds. Figures (a)–(c) illustrate the success-rate comparisons in Scenarios I, II, and III, respectively (50 trials per speed). (d) Example trajectory visualizations in Scenario II with the speed constraint of 4.0 m/s, where the flight distances (trajectory lengths) are provided. reproduce the full real world, but to emphasize features that a… view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of training dynamics over 10,000 steps. From left to right and top to bottom, the subplots show loss, success rate, maximum speed, and average speed, respectively, for the four methods. navigation under deployment-relevant shifts. Limitations. Despite these gains, performance can still degrade during high-speed flight, where reduced reaction margins amplify residual discretization and actuation … view at source ↗
read the original abstract

End-to-end unmanned aerial vehicle (UAV) navigation can achieve impressive agility in simulation, yet its obstacle-avoidance behavior often degrades after deployment because the policy must tolerate simulator mismatch, sensing irregularity, and variable-rate control. These effects are especially dangerous in cluttered environments, where stale observations or short control irregularities can directly lead to collisions. We present LNN-Fly, a deployment-oriented continuous-time navigation policy for LiDAR-based UAV obstacle avoidance. The policy combines a dynamic-programming-inspired structured recurrent update, explicit conditioning on the elapsed control interval {\Delta}t, and an input-driven adaptive forgetting gate that refreshes stale latent state near hazards while preserving consistency during sustained maneuvers. It is trained with differentiable rollouts that incorporate deployment-relevant sensing and timing perturbations. In simulation, LNN-Fly improves obstacle-avoidance performance in the tested settings and shows better tolerance to reduced control frequency, sparse observations, and control-period jitter. It also transfers zero-shot from a simplified differentiable simulator to a physical quadrotor. In indoor cross-frequency real-world tests, the system achieves 100% success over 20 flights, while policy inference has a median latency of 0.514 ms on a desktop graphics processing unit (GPU) and about 2.5 ms on the onboard central processing unit (CPU), with onboard P95 latency below 30 ms.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript introduces LNN-Fly, a continuous-time LiDAR-based UAV obstacle-avoidance policy that combines a dynamic-programming-inspired structured recurrent update, explicit conditioning on the elapsed control interval Δt, and an input-driven adaptive forgetting gate. It is trained end-to-end via differentiable rollouts that inject sensing irregularity and variable-rate control perturbations, and the authors report improved robustness to reduced control frequency, sparse observations, and jitter in simulation, plus zero-shot transfer to a physical quadrotor that achieves 100% success across 20 indoor cross-frequency flights with low inference latency.

Significance. If the central empirical claims are substantiated with adequate quantitative detail and verification that the injected perturbations match physical timing statistics, the work would address a practically important gap in deploying learned continuous-time policies on resource-constrained UAVs where control-period jitter and observation staleness are common failure modes.

major comments (2)
  1. [Abstract] Abstract: the headline claim of 100% success over 20 real-world flights is presented without any accompanying quantitative metrics (collision distance, path length, success rate versus baselines, or variance across trials), preventing assessment of effect size or comparison to prior art.
  2. [Training and evaluation sections] Training and evaluation sections: the perturbations for sensing irregularity and variable-rate control are described as 'deployment-relevant,' yet no quantitative matching (empirical CDFs, moment statistics, or Kolmogorov-Smirnov tests) is supplied between the simulated Δt and observation-age distributions and those logged from the physical platform. This directly bears on the validity of the zero-shot transfer and robustness claims.
minor comments (1)
  1. [Abstract] Abstract: the reported median latency of 0.514 ms on GPU should clarify whether this figure includes the full sensing-to-actuation pipeline or only the policy forward pass.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below and indicate where revisions will be made to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the headline claim of 100% success over 20 real-world flights is presented without any accompanying quantitative metrics (collision distance, path length, success rate versus baselines, or variance across trials), preventing assessment of effect size or comparison to prior art.

    Authors: We agree that the abstract would benefit from additional quantitative context to allow readers to assess effect size. In the revised manuscript we will expand the abstract to include median minimum collision distances, average path lengths, success rates versus the main baselines, and standard deviation across the 20 trials while retaining the 100% success statement. The detailed per-trial data already appear in the experimental section and will be cross-referenced. revision: yes

  2. Referee: [Training and evaluation sections] Training and evaluation sections: the perturbations for sensing irregularity and variable-rate control are described as 'deployment-relevant,' yet no quantitative matching (empirical CDFs, moment statistics, or Kolmogorov-Smirnov tests) is supplied between the simulated Δt and observation-age distributions and those logged from the physical platform. This directly bears on the validity of the zero-shot transfer and robustness claims.

    Authors: The observation is correct; the original submission does not supply direct distributional comparisons. We will add, in the revised training and evaluation sections, empirical CDF plots, first- and second-moment statistics, and Kolmogorov-Smirnov test p-values between the simulated perturbation distributions and the logged physical-platform timing traces. These additions will be placed immediately after the description of the perturbation model. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical results from trained policy with no derivation chain reducing to inputs

full rationale

The paper describes an end-to-end trained continuous-time policy using differentiable rollouts with injected perturbations for timing and sensing mismatch. Performance claims (simulation improvements, zero-shot transfer, 100% real-world success) are presented as experimental outcomes rather than analytical predictions or derivations. No equations, self-citations, or fitted parameters are shown that reduce reported metrics to quantities defined by construction from the inputs. The work is self-contained against external benchmarks via direct testing.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no free parameters, axioms, or invented entities are explicitly introduced or fitted in the provided text. The approach rests on standard assumptions of reinforcement learning and continuous-time control.

pith-pipeline@v0.9.1-grok · 5787 in / 1080 out tokens · 35423 ms · 2026-06-30T09:53:27.389176+00:00 · methodology

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

41 extracted references · 7 canonical work pages · 3 internal anchors

  1. [1]

    Champion-level drone racing using deep reinforce- ment learning,

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

    2023

  2. [2]

    Reaching the limit in autonomous racing: Optimal control versus reinforcement learning,

    Y . Song, A. Romero, M. M ¨uller, V . Koltun, and D. Scaramuzza, “Reaching the limit in autonomous racing: Optimal control versus reinforcement learning,”Science Robotics, vol. 8, no. 82, p. eadg1462, 2023

    2023

  3. [3]

    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

    2021

  4. [4]

    Learning vision-based agile flight via differentiable physics,

    Y . Zhang, Y . Hu, Y . Song, D. Zou, and W. Lin, “Learning vision-based agile flight via differentiable physics,”Nature Machine Intelligence, vol. 7, no. 6, pp. 954–966, 2025

    2025

  5. [5]

    Seeing through pixel motion: Learning obstacle avoidance from optical flow with one camera,

    Y . Hu, Y . Zhang, Y . Song, Y . Deng, F. Yu, L. Zhang, W. Lin, D. Zou, and W. Yu, “Seeing through pixel motion: Learning obstacle avoidance from optical flow with one camera,”IEEE Robotics and Automation Letters, 2025

    2025

  6. [6]

    A fully-autonomous aerial robot for search and rescue applications in indoor environments using learning-based techniques,

    C. Sampedro, A. Rodriguez-Ramos, H. Bavle, A. Carrio, P. De la Puente, and P. Campoy, “A fully-autonomous aerial robot for search and rescue applications in indoor environments using learning-based techniques,”Journal of Intelligent & Robotic Systems, vol. 95, no. 2, pp. 601–627, 2019

    2019

  7. [7]

    Learning to fly by myself: A self- supervised cnn-based approach for autonomous navigation,

    A. Kouris and C.-S. Bouganis, “Learning to fly by myself: A self- supervised cnn-based approach for autonomous navigation,” in2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2018, pp. 1–9

    2018

  8. [8]

    Shrinking pomcp: A framework for real-time uav search and rescue,

    Y . Zhang, B. Luo, A. Mukhopadhyay, D. Stojcsics, D. Elenius, A. Roy, S. Jha, M. Maroti, X. Koutsoukos, G. Karsaiet al., “Shrinking pomcp: A framework for real-time uav search and rescue,” in2024 International Conference on Assured Autonomy (ICAA). IEEE, 2024, pp. 48–57

    2024

  9. [9]

    Uav path planning for wildfires-sustainably fighting wildfires with automated path planning for uavs,

    Z. Hasan, S. Kumar, V . Patel, N. Poplavskyy, P. Subbaraman, and N. Xue, “Uav path planning for wildfires-sustainably fighting wildfires with automated path planning for uavs,” 2022

    2022

  10. [10]

    Crossing the reality gap: A survey on sim-to-real transferability of robot controllers in reinforcement learning,

    E. Salvato, G. Fenu, E. Medvet, and F. A. Pellegrino, “Crossing the reality gap: A survey on sim-to-real transferability of robot controllers in reinforcement learning,”IEEE Access, vol. 9, pp. 153 171–153 187, 2021

    2021

  11. [11]

    Model-free uav navigation in unknown complex environments using vision-based reinforcement learning,

    H. Wu, W. Wang, T. Wang, and S. Suzuki, “Model-free uav navigation in unknown complex environments using vision-based reinforcement learning,”Drones, vol. 9, no. 8, p. 566, 2025

    2025

  12. [12]

    Long short-term memory,

    S. Hochreiter and J. Schmidhuber, “Long short-term memory,”Neural computation, vol. 9, no. 8, pp. 1735–1780, 1997

    1997

  13. [13]

    Gaussian splatting to real world flight navigation transfer with liquid networks,

    A. Quach, M. Chahine, A. Amini, R. Hasani, and D. Rus, “Gaussian splatting to real world flight navigation transfer with liquid networks,” arXiv preprint arXiv:2406.15149, 2024

    work page arXiv 2024

  14. [14]

    Latent ordinary differential equations for irregularly-sampled time series,

    Y . Rubanova, R. T. Chen, and D. K. Duvenaud, “Latent ordinary differential equations for irregularly-sampled time series,”Advances in neural information processing systems, vol. 32, 2019

    2019

  15. [15]

    Learning on the fly: Rapid policy adaptation via differentiable simu- lation,

    J. Pan, J. Xing, R. Reiter, Y . Zhai, E. Aljalbout, and D. Scaramuzza, “Learning on the fly: Rapid policy adaptation via differentiable simu- lation,”IEEE Robotics and Automation Letters, 2026

    2026

  16. [16]

    Robust and efficient quadrotor trajectory generation for fast autonomous flight,

    B. Zhou, F. Gao, L. Wang, C. Liu, and S. Shen, “Robust and efficient quadrotor trajectory generation for fast autonomous flight,”IEEE Robotics and Automation Letters, vol. 4, no. 4, pp. 3529–3536, 2019

    2019

  17. [17]

    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 Robotics and Automation Letters, vol. 6, no. 2, pp. 478–485, 2020

    2020

  18. [18]

    Ego-swarm: A fully autonomous and decentralized quadrotor swarm system in cluttered environments,

    X. Zhou, J. Zhu, H. Zhou, C. Xu, and F. Gao, “Ego-swarm: A fully autonomous and decentralized quadrotor swarm system in cluttered environments,” in2021 IEEE international conference on robotics and automation (ICRA). IEEE, 2021, pp. 4101–4107

    2021

  19. [19]

    Toward low-flying autonomous mav trail navigation using deep neural net- works for environmental awareness,

    N. Smolyanskiy, A. Kamenev, J. Smith, and S. Birchfield, “Toward low-flying autonomous mav trail navigation using deep neural net- works for environmental awareness,” in2017 IEEE/RSJ international conference on intelligent robots and systems (IROS). IEEE, 2017, pp. 4241–4247

    2017

  20. [20]

    You only plan once: A learning-based one-stage planner with guidance learning,

    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, vol. 9, no. 7, pp. 6083–6090, 2024

    2024

  21. [21]

    Yopov2-tracker: An end-to-end agile tracking and navigation frame- work from perception to action,

    J. Lu, Y . Hui, X. Zhang, W. Feng, H. Shen, Z. Li, and B. Tian, “Yopov2-tracker: An end-to-end agile tracking and navigation frame- work from perception to action,”arXiv preprint arXiv:2505.06923, 2025

    work page arXiv 2025

  22. [22]

    Vision transformers for end-to-end vision- based quadrotor obstacle avoidance,

    A. Bhattacharya, N. Rao, D. Parikh, P. Kunapuli, Y . Wu, Y . Tao, N. Matni, and V . Kumar, “Vision transformers for end-to-end vision- based quadrotor obstacle avoidance,” in2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2025, pp. 1–8

    2025

  23. [23]

    Dream to Fly: Model-Based Reinforcement Learning for Vision-Based Drone Flight

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

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  24. [24]

    Singer: An onboard generalist vision-language navigation policy for drones,

    M. Adang, J. Low, O. Shorinwa, and M. Schwager, “Singer: An onboard generalist vision-language navigation policy for drones,” arXiv preprint arXiv:2509.18610, 2025

    work page arXiv 2025

  25. [25]

    Xu,Reasoning for Representations for Learning-Based Control

    Z. Xu,Reasoning for Representations for Learning-Based Control. University of California, Berkeley, 2021

    2021

  26. [26]

    Domain randomization for transferring deep neural networks from simulation to the real world,

    J. Tobin, R. Fong, A. Ray, J. Schneider, W. Zaremba, and P. Abbeel, “Domain randomization for transferring deep neural networks from simulation to the real world,” in2017 IEEE/RSJ international con- ference on intelligent robots and systems (IROS). IEEE, 2017, pp. 23–30

    2017

  27. [27]

    Solving Rubik's Cube with a Robot Hand

    I. Akkaya, M. Andrychowicz, M. Chociej, M. Litwin, B. McGrew, A. Petron, A. Paino, M. Plappert, G. Powell, R. Ribaset al., “Solving rubik’s cube with a robot hand,”arXiv preprint arXiv:1910.07113, 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1910

  28. [28]

    Contrastive learning for enhancing robust scene transfer in vision- based agile flight,

    J. Xing, L. Bauersfeld, Y . Song, C. Xing, and D. Scaramuzza, “Contrastive learning for enhancing robust scene transfer in vision- based agile flight,” in2024 IEEE international conference on robotics and automation (ICRA). IEEE, 2024, pp. 5330–5337

    2024

  29. [29]

    Environ- ment as policy: Learning to race in unseen tracks,

    H. Wang, J. Xing, N. Messikommer, and D. Scaramuzza, “Environ- ment as policy: Learning to race in unseen tracks,” in2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2025, pp. 11 333–11 339

    2025

  30. [30]

    Model-agnostic meta-learning for fast adaptation of deep networks,

    C. Finn, P. Abbeel, and S. Levine, “Model-agnostic meta-learning for fast adaptation of deep networks,” inInternational conference on machine learning. PMLR, 2017, pp. 1126–1135

    2017

  31. [31]

    RMA: Rapid Motor Adaptation for Legged Robots

    A. Kumar, Z. Fu, D. Pathak, and J. Malik, “Rma: Rapid motor adaptation for legged robots,”arXiv preprint arXiv:2107.04034, 2021

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  32. [32]

    Neurobem: Hybrid aerodynamic quadrotor model,

    L. Bauersfeld, E. Kaufmann, P. Foehn, S. Sun, and D. Scaramuzza, “Neurobem: Hybrid aerodynamic quadrotor model,”arXiv preprint arXiv:2106.08015, 2021

    work page arXiv 2021

  33. [33]

    Safe reinforcement learning via shielding,

    M. Alshiekh, R. Bloem, R. Ehlers, B. K ¨onighofer, S. Niekum, and U. Topcu, “Safe reinforcement learning via shielding,” inProceedings of the AAAI conference on artificial intelligence, vol. 32, no. 1, 2018

    2018

  34. [34]

    Navrl: Learning safe flight in dynamic environments,

    Z. Xu, X. Han, H. Shen, H. Jin, and K. Shimada, “Navrl: Learning safe flight in dynamic environments,”IEEE Robotics and Automation Letters, 2025

    2025

  35. [35]

    Autonomous quadrotor obstacle avoidance based on dueling double deep recurrent q-learning with monocular vision,

    J. Ou, X. Guo, M. Zhu, and W. Lou, “Autonomous quadrotor obstacle avoidance based on dueling double deep recurrent q-learning with monocular vision,”Neurocomputing, vol. 441, pp. 300–310, 2021

    2021

  36. [36]

    Neu- ral ordinary differential equations,

    R. T. Chen, Y . Rubanova, J. Bettencourt, and D. K. Duvenaud, “Neu- ral ordinary differential equations,”Advances in neural information processing systems, vol. 31, 2018

    2018

  37. [37]

    Neural controlled differential equations for irregular time series,

    P. Kidger, J. Morrill, J. Foster, and T. Lyons, “Neural controlled differential equations for irregular time series,”Advances in neural information processing systems, vol. 33, pp. 6696–6707, 2020

    2020

  38. [38]

    Liquid time-constant networks,

    R. Hasani, M. Lechner, A. Amini, D. Rus, and R. Grosu, “Liquid time-constant networks,” inProceedings of the AAAI conference on artificial intelligence, vol. 35, no. 9, 2021, pp. 7657–7666

    2021

  39. [39]

    Closed-form continuous- time neural networks,

    R. Hasani, M. Lechner, A. Amini, L. Liebenwein, A. Ray, M. Tschaikowski, G. Teschl, and D. Rus, “Closed-form continuous- time neural networks,”Nature Machine Intelligence, vol. 4, no. 11, pp. 992–1003, 2022

    2022

  40. [40]

    Robust flight navigation with liquid neural networks,

    P. Kao, “Robust flight navigation with liquid neural networks,” Ph.D. dissertation, Massachusetts Institute of Technology, 2022

    2022

  41. [41]

    Flying on point clouds with reinforcement learning,

    G. Xu, T. Wu, Z. Wang, Q. Wang, and F. Gao, “Flying on point clouds with reinforcement learning,” in2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2025, pp. 7231–7238

    2025