AEOS: Active Environment-aware Optimal Scanning Control for UAV LiDAR-Inertial Odometry in Complex Scenes

Jianping Li; Lihua Xie; Muqing Cao; Shenghai Yuan; Xinhang Xu; Zhongyuan Liu

REVIEW 5 major objections 5 minor 43 references

AEOS: Active Environment-aware Optimal Scanning Control for UAV LiDAR-Inertial Odometry in Complex Scenes

T0 review · 5 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a hybrid MPC/RL controller that continuously re-steers a single motorized LiDAR, instead of spinning it at a fixed rate, can cut UAV LiDAR-inertial odometry error in occluded, feature-sparse scenes while running in un

desk verdict AEOS is a genuinely new hybrid MPC/RL scanning controller with a coherent architecture, but the evaluation as reported undercuts the abstract's claims: no held-out split, no fully learned baseline, no error bars, and only neural-inference timing. read the letter →

arxiv 2509.09141 v1 pith:3CUOA5ZE submitted 2025-09-11 cs.RO

Jianping Li , Xinhang Xu , Zhongyuan Liu , Shenghai Yuan , Muqing Cao , Lihua Xie This is my paper

classification cs.RO MSC 68T4093C85

keywords activesensingLiDAR-InertialOdometrymodelpredictivecontrolreinforcementlearningSLAMsensorUAVpointcloudsimulation

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

AEOS asks whether a UAV's single LiDAR can be made to act like an owl's head: continuously re-pointing itself toward whatever the environment and the vehicle's own motion make most informative. The paper's claim is that this active steering, split between an analytical uncertainty model for exploitation and a lightweight learned cost map for exploration, beats fixed-speed rotation, optimization-only control, and fully learned policies. The authors report this across simulated tunnels, caves, forests, and real industrial and basement scenes, with an average absolute pose error of 3.83 meters in simulation and 0.51 meters in the real-world tests, and with a control inference time of 0.047 milliseconds on an onboard edge unit. If true, the payoff is that task-aware, adaptive sensing can improve a core autonomy capability without adding sensors or exceeding tight payload and compute budgets.

What carries the argument

The load-bearing mechanism is a differentiable hybrid RL-MPC: the MPC layer solves a finite-horizon optimal control problem over LiDAR angular velocities, with the total cost equal to an analytically computed exploitation term plus a neural-network-generated exploration cost plus a smoothness penalty. The exploitation term is a piecewise linear, differentiable surrogate of the trace of the inverse of a surrogate Fisher information matrix built from point-to-plane residuals and Jacobians of raycast synthetic scans. The exploration term is a structured quadratic function over angle and angular velocity, whose positive-semidefinite diagonal curvature matrix is output by a small MLP from a panor

What would settle it

Retrain the AEOS policy on the first 40% of each of the eight simulation sequences, then recompute APE only over the remaining 60% of each trajectory. If the held-out APE no longer clearly beats the optimization-only and fixed-rate baselines, the reported improvement is substantially memorization rather than generalization. Alternatively, train on seven scenes and test on the eighth, reporting the leave-one-scene-out average APE.

Watch

Extended reading notes

Core claim

The central claim is that a differential hybrid controller—Model Predictive Control whose cost is the sum of an analytical predictive LiDAR observability term and a reinforcement-learned quadratic cost map—can continuously modulate the yaw angle and angular velocity of a front-mounted motorized LiDAR to improve LiDAR-inertial odometry accuracy in complex, occluded environments. The analytical term, derived from a surrogate Fisher information matrix formed by point-to-plane Jacobians of raycast synthetic scans, is made differentiable by piecewise linear interpolation so that gradients flow through the MPC layer. The learned term, a two-layer network fed by a low-resolution panoramic depth map

Load-bearing premise

The generalization claim depends on APE being measured on trajectory sections that were not used to train the policy; the paper trains on the first 40% of each sequence and then reports full-sequence APE without specifying a held-out test split.

Editorial extensions

If this is right

Adaptive, scene-aware scanning could become a drop-in corrective layer for single-LiDAR UAV odometry, narrowing the gap between compact sensors and larger multi-sensor perception systems.
The reported real-time inference (0.047 ms average, 95% under 0.059 ms on an Intel-N305 edge unit) implies the approach leaves substantial compute headroom for concurrent LIO, mapping, and planning on resource-constrained platforms.
The hybrid architecture offers an interpretable role separation: the learned cost sets a low-frequency scanning strategy while the MPC supplies high-frequency observability corrections, which may make failures easier to diagnose than in end-to-end learned control.
The point-cloud-based simulation environment, built from real-world trajectories and dense global maps, could serve as a reusable benchmark for other active LiDAR control methods.
The results suggest that replacing fixed-speed motorized scanning with even a moderately informed adaptive policy can yield outsized gains in degenerate geometries such as tunnels and industrial equipment.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A testable extension is to evaluate the policy on trajectories entirely disjoint from the 40% used for training; the full-sequence APE numbers in the paper do not demonstrate whether the benefit generalizes to unseen geometry or partly reflects memorization of the trained corridor.
The convex quadratic restriction on the learned cost, while computationally convenient, may unnecessarily limit the expressive range of exploration strategies; a larger-capacity cost network could improve performance at some inference cost, an ablation the paper does not run.
Because the analytical uncertainty term already encodes local observability, a natural next step is to couple AEOS with active trajectory planning rather than fixing the flight path, potentially compounding the odometry gains with path-level information gathering.
The owl-gaze analogy suggests a further adaptation: incorporating roll and pitch of the motor axis, not just yaw, could extend the controller's effectiveness in environments with vertical features, a direction the paper mentions only implicitly.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Brief take: the architecture is real. Combining a differentiable MPC layer with a learned quadratic cost map for motorized LiDAR scanning is a new combination in the active-SLAM literature, and the design is sensible: analytical uncertainty for exploitation, learned cost for exploration. The math in Section 4.2 is coherent — the implicit function theorem derivation is standard and correctly applied. The point-cloud simulation environment built from real-world maps is a useful contribution in itself, and the two real-world flights show the system works on hardware.

The soft spots are all in the evaluation, and they are load-bearing. The abstract claims to beat "fully learned baselines," but no fully learned baseline appears in Tables 2 or 4. "AEOS w/o J_unc" is not a fully learned policy. Second, Section 5.3 says "We select only 40% length of the trajectory for each dataset as training data," and Section 6.2 reports APE on full sequences with no held-out test split. If the reported numbers include the training portion, the gains could reflect memorization rather than generalization. The reward in Eq. (26) uses RTE, the same error family as the APE metric, so the circularity is real. Third, there are no error bars or repeated trials anywhere, so "significantly improves" is not backed by statistics. Fourth, the real-time claim is based on the neural network inference time alone (Fig. 13), excluding the MPC solve and rendering cost.

I also noticed an inconsistency in the ablation: AEOS w/o J_unc is worse than Optimization-Based MPC in three of the eight scenes (Seq03, Seq04, Seq08). That undercuts the claim that the learned exploration component consistently helps.

These issues are addressable, not fatal. The idea has merit, the system is built, and the real-world results are encouraging even if limited to two scenes. The paper needs a proper train/test split, a real fully learned baseline, repeated runs with variance, full-loop latency, and released code. With those, it could be a solid contribution.

I would send this to peer review with the expectation of major revision. The authors should be given the chance to fix the evaluation. It's not a desk reject, but the current version overclaims.

Referee Report

5 major / 5 minor

Summary. The paper proposes AEOS, a hybrid MPC/RL framework for actively controlling the yaw of a motorized LiDAR on a UAV to improve LiDAR-inertial odometry (LIO) accuracy. Exploitation of feature-rich or observable regions is handled by an analytical observability-based uncertainty cost, while exploration is guided by a lightweight neural network that outputs a quadratic cost over scan angle and angular velocity. The policy is trained with PPO through a differentiable MPC layer using an implicit-function-theorem gradient (Theorem 1). A point-cloud-based simulation environment built from real-world LiDAR maps is used for training and evaluation. Experiments report APE improvements over fixed-rate scanning, random control, an optimization-only MPC baseline, and an ablation (AEOS without the uncertainty cost) in eight simulated scenes and two real-world scenes, together with a low inference latency on an Intel-N305 edge unit.

Significance. If the results hold, the AEOS architecture is a worthwhile contribution: it combines an interpretable, model-based exploitation signal with a low-capacity learned exploration cost, and the differentiable-MPC derivation is coherent. The point-cloud simulation built from real maps and the real-world deployment on a UAV are concrete strengths, and the reported inference latency suggests the scheme is lightweight enough for onboard use. However, the paper's central comparative claim is currently not supported by the evidence as reported: the simulation evaluation may include the training portion of each trajectory, no fully learned baseline is implemented, and no variance or repeated trials are reported. These issues directly affect the headline claims of generalization and superiority over learning-based methods.

major comments (5)

[§5.3, §6.2, Eq. (26)] The training/evaluation split is not stated. Section 5.3 says 'We select only 40% length of the trajectory for each dataset as training data. In each episode, 60 second-length is extracted from the training.' Section 6.2 reports APE 'for each scene independently' over full sequences without defining a held-out test portion. If the reported Table 2 APE includes the 40% used for training, the learned component's reward (Eq. 26, r_lio = 1/RTE) optimizes the same error family as the evaluation metric, so part of the improvement could be memorization rather than generalization. The authors must specify a disjoint held-out split and report APE separately on training and test portions.
[§6.1, Tables 2 and 4, Abstract] The abstract claims AEOS 'significantly improves odometry accuracy compared to fixed-rate, optimization-only, and fully learned baselines,' and Section 8 repeats 'purely learning-based baselines.' However, Section 6.1 enumerates only Fixed-Rate (Slow/Fast), Random Control, Optimization-Based MPC, and AEOS w/o U_unc. None of these is a fully learned or end-to-end RL baseline. Tables 2 and 4 therefore do not substantiate this headline comparison. Either add a comparable end-to-end RL policy or revise the claim to match the implemented baselines.
[§6.2, Tables 2 and 4] All APE values are reported as single point estimates with no error bars, number of repeated runs, or statistical tests. This matters because Random Control is stochastic by definition, PPO training has randomness across seeds, and the real-world experiments are single trials. The word 'significantly' is therefore not justified formally. Report mean ± standard deviation over multiple runs/seeds for every method and scene, and, if possible, a paired significance test.
[Theorem 1 vs. §4.3.2, Eqs. (10)-(11), (18)] Theorem 1 assumes the MPC objective J is twice continuously differentiable in both ω and φ. The uncertainty surrogate U_unc in Eq. (18) is piecewise linear in θ, hence only piecewise differentiable and not C² on the sample grid. The paper notes 'differentiable almost everywhere' but does not reconcile this with the strict regularity condition used in the implicit function theorem. This is load-bearing for the policy-gradient derivation. A smooth surrogate (e.g., cubic spline) or a nonsmooth/implicit-differentiation extension is needed.
[§6.4] The reported real-time latency (average 0.047 ms, 95% ≤ 0.059 ms) appears to cover only the neural network inference, not the full AEOS control loop. The controller also performs uncertainty sampling via raycasting, cache construction of {U_s}, and the MPC solve. If the timing excludes these components, the real-time claim is incomplete. Report the end-to-end control-period latency, including the analytical uncertainty evaluation and MPC optimization, on the same edge hardware.

minor comments (5)

[Table 1] Simu-Seq04 is internally inconsistent: average velocity 2.40 m/s over 379 s implies a trajectory length of about 910 m, not the listed 158 m.
[Table 2] The Simu-Seq01 entry for 'AEOS w/o U_unc' and 'AEOS (Ours)' is typeset as '1.390.89'; this should be '1.39' and '0.89'.
[§4.2.1, Eq. (5)] The neural cost J_nn depends on the current observation o_t, but Eq. (5) writes J_nn(θ_k, ω_k; φ) without o_t. Clarify the notation.
[§4.4.1] Typo: 'downspample' should be 'downsample'. Also, the phrase 'Both of Q_k and q_k are conditioned' is ungrammatical.
[§6.3] The text says 'The ATE and trajectories ... are illustrated in Fig. 12,' but the tables and evaluation use APE, not ATE. Align the terminology.

Circularity Check

1 steps flagged · score 5.0 of 10

Simulation APE is reported on full trajectories while 40% of each trajectory is used for training; without a stated held-out split, the reported improvement is partly forced by the training reward.

fitted input called prediction [Section 5.3 (Training Configuration), Section 4.4.2 (Reward Functions), Section 6.2 (Evaluation in Simulation)]
""We select only 40% length of the trajectory for each dataset as training data." ... "r_lio(t)=1/RTE(t,τ), where RTE(t,τ) denotes the average relative translational error over a sliding window of duration τ" ... "Table 2 reports the Absolute Pose Error (APE) across eight simulated sequences.""

The policy is trained on 40% of each trajectory with a reward that is the reciprocal of the LIO trajectory-error measure; APE, the reported metric, is the same error family and is computed for the full sequence with no stated held-out split. For the trained 40% of each sequence, the policy was explicitly optimized to reduce RTE, so any APE improvement on those segments is a direct consequence of the training objective rather than a prediction. Since Table 2 is the main evidence for the abstract's accuracy claim, the simulation comparison is at least partially forced by construction.

full rationale

The core derivation chain is not circular in the sense of an equation reducing to its own inputs: the hybrid MPC/RL architecture is implemented, the uncertainty model and learned cost map are specified, and the system is evaluated against external baselines. The real-world experiments (Real-Seq01/02) are not used in training and therefore provide some independent support for the claimed improvement. However, the simulation evaluation has a significant train/eval overlap: 40% of each trajectory is used for training, and the reported APE in Table 2 is on full sequences without a stated held-out split. The training reward r_lio = 1/RTE is the same error family as the evaluation APE, so for the trained portions the improvement is memorization/optimization rather than generalization. This raises the circularity score to 5. The self-citation to UA-MPC (Li et al., 2025) in Section 4.3 is provenance rather than load-bearing circularity, because the equations are reproduced in the paper and the method is implemented. The advertised comparison against a 'fully learned baseline' is missing from Tables 2 and 4, but that is a completeness/correctness issue, not circularity.

Assumptions & free parameters 7 free parameters · 7 assumptions · 0 invented entities

The central method rests on hand-chosen hyperparameters and domain assumptions about observability, interpolation, simulation fidelity, and evaluation integrity. No new physical entities are introduced; the learned cost map and panoramic representation are algorithmic constructs.

free parameters (7)

PPO policy network parameters φ = trained over 500k environment steps per scene
Central to the learned exploration cost J_nn; fitted via a reward that includes LIO error.
Reward weights λ_exp and λ_lio = not reported
Balance exploration and localization in Eq. (24); no numeric values are given, so the training objective is not fully specified.
MPC horizon T and control interval Δt = not reported
Define the planning horizon and discrete-time dynamics in Eqs. (2) and (5); hand-chosen and needed to reproduce the controller.
Uncertainty surrogate sampling interval Δθ = not reported
Controls the number of precomputed U(θ) samples (N = 2π/Δθ) and the quality of the piecewise linear approximation in Eq. (18).
Neural cost bounds q_min and q_max = 0.1 and 1e5
Hand-chosen scaling limits in Eq. (23) for the learned quadratic cost; affect exploration aggressiveness.
Training data fraction per trajectory = 40% of each trajectory
Chosen in Section 5.3; if the evaluation sequence overlaps this training data, reported APE is partially fitted.
PPO hyperparameters = KL threshold 0.01, gradient clip 0.5
Part of the training recipe; hand-chosen and not justified by sensitivity analysis.

assumptions (7)

domain assumption First-order motor dynamics θ_{t+1} = θ_t + ω_t Δt adequately model the rotating LiDAR
Section 4.1.1 explicitly neglects higher-order rotational dynamics; this is load-bearing for MPC prediction.
domain assumption Trace of the inverse Fisher information from point-to-plane Jacobians predicts future LIO pose uncertainty
Section 4.3.1 applies A-optimal design to rendered LiDAR points; assumes observability is the dominant driver of odometry error.
domain assumption Piecewise linear interpolation of U(θ) is a valid differentiable surrogate because adjacent scan angles yield similar uncertainty
Remark 1 justifies the surrogate used in Eq. (18); it is asserted, not proven.
domain assumption Raycasting from the panoramic depth map produces synthetic scans that preserve the observability structure of the real LiDAR
Used in both the uncertainty model (Section 4.3) and the simulation environment (Section 5.1); sim-to-real validity depends on it.
ad hoc to paper The MPC objective J is twice continuously differentiable as required by Theorem 1
Theorem 1 invokes standard regularity, but J_unc is only piecewise differentiable, so the assumption is violated at interpolation knots.
domain assumption The point-cloud simulation with expert trajectories and Fast-LIO2 transfers to real-world LIO performance
The sim-to-real transfer claim in Sections 5 and 6.3 presupposes that the raycast environment faithfully reproduces real LIO failure modes.
domain assumption The real-world prior map is used only for ground truth and safety, not as an input to the LIO backend whose APE is reported
Section 6.3 says the prior map enables 'robust pose estimation'; if LIO is augmented with map registration, the APE no longer measures standalone LIO accuracy.

how reviews work

0 comments

Cite this review

Pith. "Pith review of AEOS: Active Environment-aware Optimal Scanning Control for UAV LiDAR-Inertial Odometry in Complex Scenes." pith.science (2026). https://pith.science/paper/3CUOA5ZE

@misc{pith2026250909141,
  author       = {Pith},
  title        = {Pith review of: AEOS: Active Environment-aware Optimal Scanning Control for UAV LiDAR-Inertial Odometry in Complex Scenes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3CUOA5ZE}},
  note         = {Machine review of arXiv:2509.09141}
}

read the original abstract

LiDAR-based 3D perception and localization on unmanned aerial vehicles (UAVs) are fundamentally limited by the narrow field of view (FoV) of compact LiDAR sensors and the payload constraints that preclude multi-sensor configurations. Traditional motorized scanning systems with fixed-speed rotations lack scene awareness and task-level adaptability, leading to degraded odometry and mapping performance in complex, occluded environments. Inspired by the active sensing behavior of owls, we propose AEOS (Active Environment-aware Optimal Scanning), a biologically inspired and computationally efficient framework for adaptive LiDAR control in UAV-based LiDAR-Inertial Odometry (LIO). AEOS combines model predictive control (MPC) and reinforcement learning (RL) in a hybrid architecture: an analytical uncertainty model predicts future pose observability for exploitation, while a lightweight neural network learns an implicit cost map from panoramic depth representations to guide exploration. To support scalable training and generalization, we develop a point cloud-based simulation environment with real-world LiDAR maps across diverse scenes, enabling sim-to-real transfer. Extensive experiments in both simulation and real-world environments demonstrate that AEOS significantly improves odometry accuracy compared to fixed-rate, optimization-only, and fully learned baselines, while maintaining real-time performance under onboard computational constraints. The project page can be found at https://kafeiyin00.github.io/AEOS/.

Figures

Figures reproduced from arXiv: 2509.09141 by the authors.

**Figure 1.** Overview of the proposed Active Environment-aware Optimal Scanning Control (AEOS). (a) We proposed a bio-inspired active panoramic sensing UAV system for complex scenes. (b) The active sensing agent practices and learns knowledge in a point cloud-based simulation environment via reinforcement learning. (c) The proposed agent is a differentiable hybrid Model Predictive Control (MPC) considering implicit information f… view at source ↗

**Figure 2.** Hardware design of the proposed Active Environment-aware Optimal Scanning (AEOS)-Drone. (a) Sensor synchronization and embedded system architecture. (b) Bio-inspired UAV hardware design and related coordination systems. (c) Coordination systems related to the AEOS-Drone. 3.3. Coordination Systems and Calibration To ensure accurate transformation between sensor data and global pose estimates, AEOS defines a set of co… view at source ↗

**Figure 3.** Unified Panoramic Representation of Local Point Clouds for Downstream Modules. uncertainty modeling (Section 4.3), while a coarser resolution suffices for neural cost learning (Section 4.4) to reduce computational burden. This design promotes representational consistency across modules and reduces redundant computation during real-time onboard execution. The analytical uncertainty model (detailed in Section 4.3) p… view at source ↗

Figures from the paper (10 more)

**Figure 4.** Figure 4: Scanning prediction in a control horizon using the unified panoramic depth map. where [⋅]× denotes the standard skew-symmetric operator mapping ℝ3 → ℝ3×3 . According to A-optimal design theory Pukelsheim (2006), the expected localization uncertainty under a candidate L…

**Figure 6.** Figure 6: System architecture of the point cloud-based simulation environment. Ground-truth data are used to simulate LiDAR observations and UAV dynamics, enabling reinforcement learning with real-time feedback from the LIO backend. timestep 𝑡, the total reward is defined as: 𝑟…

**Figure 7.** Figure 7: Overview of the simulation dataset used for active scanning. (a) High-fidelity point cloud maps of diverse real-world environments, including caves, tunnels, subways, forests, and lava tubes across the globe. (b) The simulation framework integrates the PX4 autopilot, …

**Figure 8.** Figure 8: Simulation experiments. (a)-(g), the screenshots of the simulation environments for the sequences Simu-Seq01 to Simu-Seq08. Each figure includes the point clouds, panoramic depth, rotor angle serial, and rotor speed serial [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]

**Figure 9.** Figure 9: Simulation evaluations of the proposed AEOS. (a)-(g), the trajectories and ATE of the proposed AEOS for Simu-Seq01 to SimuSeq08. : Preprint submitted to Elsevier Page 14 of 18 [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]

**Figure 10.** Figure 10: Real-world evaluation of the proposed AEOS in real-seq1 [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]

**Figure 11.** Figure 11: Real-world evaluation of the proposed AEOS in real-seq2 [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]

**Figure 12.** Figure 12: Real-world evaluation.(a) and (b) the trajectories and ATE of the proposed AEOS for Real-Seq01 to Real-Seq02. : Preprint submitted to Elsevier Page 15 of 18 [PITH_FULL_IMAGE:figures/full_fig_p015_12.png]

**Figure 13.** Figure 13: Inference time distribution of the proposed AEOS control policy evaluated on an x86 Intel-N305 edge computing unit. The results show that the system achieves real-time performance with low latency, suitable for onboard deployment. interpretability and task-awareness v…

**Figure 15.** Figure 15: Application for rescue in complex scenes. (a) Flying through the window. (b) Flying through the narrow corridor. (c) Flying through the door. (d) Quick exploration of the complex environment. The corresponding video can be found on the project page. corridors, thereby…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 7 linked inside Pith

[1]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in ":" * " " * FUNCTION f...
[2]

, author Bernreiter, L

author Ebadi, K. , author Bernreiter, L. , author Biggie, H. , author Catt, G. , author Chang, Y. , author Chatterjee, A. , author Denniston, C.E. , author Desch \^e nes, S.P. , author Harlow, K. , author Khattak, S. , et al., year 2023 . title Present and future of slam in extreme environments: The darpa subt challenge . journal IEEE Transactions on Robo...

2023
[3]

, author Wu, W

author Li, J. , author Wu, W. , author Yang, B. , author Zou, X. , author Yang, Y. , author Zhao, X. , author Dong, Z. , year 2023 . title Whu-helmet: A helmet-based multisensor slam dataset for the evaluation of real-time 3-d mapping in large-scale gnss-denied environments . journal IEEE Transactions on Geoscience and Remote Sensing volume 61 , pages 1--16

2023
[4]

, author Huang, H

author Chen, X. , author Huang, H. , author Liu, Y. , author Li, J. , author Liu, M. , year 2022 . title Robot for automatic waste sorting on construction sites . journal Automation in Construction volume 141 , pages 104387

2022
[5]

, author Yuan, S

author Li, J. , author Yuan, S. , author Cao, M. , author Nguyen, T.M. , author Cao, K. , author Xie, L. , year 2024 . title Hcto: Optimality-aware lidar inertial odometry with hybrid continuous time optimization for compact wearable mapping system . journal ISPRS Journal of Photogrammetry and Remote Sensing volume 211 , pages 228--243

2024
[6]

, author Yang, B

author Li, J. , author Yang, B. , author Cong, Y. , author Cao, L. , author Fu, X. , author Dong, Z. , year 2019 . title 3d forest mapping using a low-cost uav laser scanning system: Investigation and comparison . journal Remote Sensing volume 11 , pages 717

2019
[7]

, author Ramezani, M

author Proudman, A. , author Ramezani, M. , author Digumarti, S.T. , author Chebrolu, N. , author Fallon, M. , year 2022 . title Towards real-time forest inventory using handheld lidar . journal Robotics and Autonomous Systems volume 157 , pages 104240

2022
[8]

, author Wang, W

author Wang, G. , author Wang, W. , author Ding, P. , author Liu, Y. , author Wang, H. , author Fan, Z. , author Bai, H. , author Hongbiao, Z. , author Du, Z. , year 2023 . title Development of a search and rescue robot system for the underground building environment . journal Journal of field robotics volume 40 , pages 655--683

2023

Show all 43 references

[9]

, author Kong, F

author Chen, N. , author Kong, F. , author Xu, W. , author Cai, Y. , author Li, H. , author He, D. , author Qin, Y. , author Zhang, F. , year 2023 . title A self-rotating, single-actuated uav with extended sensor field of view for autonomous navigation . journal Science Roboti...

2023
[10]

, author Vlaminck, M

author Diels, L. , author Vlaminck, M. , author De Wit, B. , author Philips, W. , author Luong, H. , year 2022 . title On the optimal mounting angle for a spinning lidar on a uav . journal IEEE Sensors Journal volume 22 , pages 21240--21247

2022
[11]

, author Wang, D

author Chen, Y. , author Wang, D. , author Thomas, L. , author Dantu, K. , author Koppal, S.J. , year 2024 . title Design of an adaptive lightweight lidar to decouple robot-camera geometry . journal IEEE Transactions on Robotics

2024
[12]

, author Zlot, R

author Kaul, L. , author Zlot, R. , author Bosse, M. , year 2016 . title Continuous-time three-dimensional mapping for micro aerial vehicles with a passively actuated rotating laser scanner . journal Journal of Field Robotics volume 33 , pages 103--132

2016
[13]

, author Browning, B

author Alismail, H. , author Browning, B. , year 2015 . title Automatic calibration of spinning actuated lidar internal parameters . journal Journal of Field Robotics volume 32 , pages 723--747

2015
[14]

, author He, Y

author Cui, J. , author He, Y. , author Niu, J. , author Ouyang, Z. , author Xing, G. , year 2024 . title lidar: An adaptive high-resolution panoramic lidar system , in: booktitle Proceedings of the 30th Annual International Conference on Mobile Computing and Networking , pp. ...

2024
[15]

, author Xu, X

author Li, J. , author Xu, X. , author Liu, J. , author Cao, K. , author Yuan, S. , author Xie, L. , year 2025 . title Ua-mpc: Uncertainty-aware model predictive control for motorized lidar odometry . journal IEEE Robotics and Automation Letters

2025
[16]

, author Huang, S

author Leung, C. , author Huang, S. , author Kwok, N. , author Dissanayake, G. , year 2006 . title Planning under uncertainty using model predictive control for information gathering . journal Robotics and Autonomous Systems volume 54 , pages 898--910

2006
[17]

, author Song, Y

author Romero, A. , author Song, Y. , author Scaramuzza, D. , year 2024 . title Actor-critic model predictive control , in: booktitle 2024 IEEE International Conference on Robotics and Automation (ICRA) , organization IEEE . pp. pages 14777--14784

2024
[18]

, author Nandha, D

author Morgan, A.S. , author Nandha, D. , author Chalvatzaki, G. , author D’Eramo, C. , author Dollar, A.M. , author Peters, J. , year 2021 . title Model predictive actor-critic: Accelerating robot skill acquisition with deep reinforcement learning , in: booktitle 2021 IEEE In...

2021
[19]

, author Li, J

author Zhang, Y. , author Li, J. , author Zhao, X. , author Liang, F. , author Dong, Z. , author Yang, B. , year 2025 . title Armor: Adaptive meshing with reinforcement optimization for real-time 3d monitoring in unexposed scenes . journal arXiv preprint arXiv:2504.19624

2025 arXiv
[20]

, author Yao, Y

author Ling, Z. , author Yao, Y. , author Li, X. , author Su, H. , year 2023 . title On the efficacy of 3d point cloud reinforcement learning . journal arXiv preprint arXiv:2306.06799

2023 arXiv
[21]

, author Xu, X

author Cao, M. , author Xu, X. , author Yang, Y. , author Li, J. , author Jin, T. , author Wang, P. , author Hung, T.Y. , author Lin, G. , author Xie, L. , year 2024 . title Learning dynamic weight adjustment for spatial-temporal trajectory planning in crowd navigation . journ...

2024 arXiv
[22]

, author Chen, H

author Wang, Z. , author Chen, H. , author Zhang, S. , author Lou, Y. , year 2023 . title Active view planning for visual slam in outdoor environments based on continuous information modeling . journal IEEE/ASME Transactions On Mechatronics volume 29 , pages 237--248

2023
[23]

, author Lai, G

author Shi, C. , author Lai, G. , author Yu, Y. , author Bellone, M. , author Lippiello, V. , year 2023 . title Real-time multi-modal active vision for object detection on uavs equipped with limited field of view lidar and camera . journal IEEE Robotics and Automation Letters ...

2023
[24]

, author Xie, L

author Cao, K. , author Xie, L. , year 2023 . title Trust-region inverse reinforcement learning . journal IEEE Transactions on Automatic Control volume 69 , pages 1037--1044

2023
[25]

, author Teixeira, L

author Bartolomei, L. , author Teixeira, L. , author Chli, M. , year 2021 . title Semantic-aware active perception for uavs using deep reinforcement learning , in: booktitle 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , organization IEEE . p...

2021
[26]

, author Singh, S

author Zhang, J. , author Singh, S. , et al., year 2014 . title Loam: Lidar odometry and mapping in real-time. , in: booktitle Robotics: Science and systems , organization Berkeley, CA . pp. pages 1--9

2014
[27]

, author Liu, Z

author Li, J. , author Liu, Z. , author Xu, X. , author Liu, J. , author Yuan, S. , author Xu, F. , author Xie, L. , year 2025 . title Limo-calib: On-site fast lidar-motor calibration for quadruped robot-based panoramic 3d sensing system . journal arXiv preprint arXiv:2502.12655

2025 arXiv
[28]

, author Wu, K

author Chen, J. , author Wu, K. , author Hu, M. , author Suganthan, P.N. , author Makur, A. , year 2024 . title Lidar-based end-to-end active slam using deep reinforcement learning in large-scale environments . journal IEEE Transactions on Vehicular Technology

2024
[29]

, author Khosoussi, K

author Ramezani, M. , author Khosoussi, K. , author Catt, G. , author Moghadam, P. , author Williams, J. , author Borges, P. , author Pauling, F. , author Kottege, N. , year 2022 . title Wildcat: Online continuous-time 3d lidar-inertial slam . journal arXiv preprint arXiv:2205.12595

2022 arXiv
[30]

, author Goldschmid, P

author Bonetto, E. , author Goldschmid, P. , author Pabst, M. , author Black, M.J. , author Ahmad, A. , year 2022 . title irotate: Active visual slam for omnidirectional robots . journal Robotics and Autonomous Systems volume 154 , pages 104102

2022
[31]

, author Rodr \' guez, J.J.G

author Placed, J.A. , author Rodr \' guez, J.J.G. , author Tard \'o s, J.D. , author Castellanos, J.A. , year 2022 . title Explorb-slam: Active visual slam exploiting the pose-graph topology , in: booktitle Iberian Robotics conference , organization Springer . pp. pages 199--210

2022
[32]

, author Strader, J

author Placed, J.A. , author Strader, J. , author Carrillo, H. , author Atanasov, N. , author Indelman, V. , author Carlone, L. , author Castellanos, J.A. , year 2023 . title A survey on active simultaneous localization and mapping: State of the art and new frontiers . journal...

2023
[33]

, author Hahnel, D

author Stachniss, C. , author Hahnel, D. , author Burgard, W. , year 2004 . title Exploration with active loop-closing for fastslam , in: booktitle 2004 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)(IEEE Cat. No. 04CH37566) , organization IEEE . pp...

2004
[34]

, author Guo, H

author Bai, R. , author Guo, H. , author Yau, W.Y. , author Xie, L. , year 2024 . title Graph-based slam-aware exploration with prior topo-metric information . journal IEEE Robotics and Automation Letters

2024
[35]

, author Yang, C.y

author Li, P. , author Yang, C.y. , author Wang, R. , author Wang, S. , year 2020 . title A high-efficiency, information-based exploration path planning method for active simultaneous localization and mapping . journal International Journal of Advanced Robotic Systems volume 1...

2020
[36]

, author Schlotfeldt, B

author Kantaros, Y. , author Schlotfeldt, B. , author Atanasov, N. , author Pappas, G.J. , year 2019 . title Asymptotically optimal planning for non-myopic multi-robot information gathering. , in: booktitle Robotics: Science and Systems , pp. pages 22--26

2019
[37]

, author Malobick \`y , B

author Mih \'a lik, M. , author Malobick \`y , B. , author Peniak, P. , author Vestenick \`y , P. , year 2022 . title The new method of active slam for mapping using lidar . journal Electronics volume 11 , pages 1082

2022
[38]

, author Gandhi, D

author Chaplot, D.S. , author Gandhi, D. , author Gupta, S. , author Gupta, A. , author Salakhutdinov, R. , year 2020 . title Learning to explore using active neural slam . journal arXiv preprint arXiv:2004.05155

2020 arXiv
[39]

, author Wu, T

author Xu, G. , author Wu, T. , author Wang, Z. , author Wang, Q. , author Gao, F. , year 2025 a. title Flying on point clouds with reinforcement learning . journal arXiv preprint arXiv:2503.00496

2025 arXiv
[40]

, author Han, X

author Xu, Z. , author Han, X. , author Shen, H. , author Jin, H. , author Shimada, K. , year 2025 b. title Navrl: Learning safe flight in dynamic environments . journal IEEE Robotics and Automation Letters

2025
[41]

, author Xu, J

author Lv, J. , author Xu, J. , author Hu, K. , author Liu, Y. , author Zuo, X. , year 2020 . title Targetless calibration of lidar-imu system based on continuous-time batch estimation , in: booktitle 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IR...

2020
[42]

, year 2006

author Pukelsheim, F. , year 2006 . title Optimal design of experiments . publisher SIAM

2006
[43]

, author Cai, Y

author Xu, W. , author Cai, Y. , author He, D. , author Lin, J. , author Zhang, F. , year 2022 . title Fast-lio2: Fast direct lidar-inertial odometry . journal IEEE Transactions on Robotics volume 38 , pages 2053--2073

2022

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [1]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in ":" * " " * FUNCTION f...

[2] [2]

, author Bernreiter, L

author Ebadi, K. , author Bernreiter, L. , author Biggie, H. , author Catt, G. , author Chang, Y. , author Chatterjee, A. , author Denniston, C.E. , author Desch \^e nes, S.P. , author Harlow, K. , author Khattak, S. , et al., year 2023 . title Present and future of slam in extreme environments: The darpa subt challenge . journal IEEE Transactions on Robo...

2023

[3] [3]

, author Wu, W

author Li, J. , author Wu, W. , author Yang, B. , author Zou, X. , author Yang, Y. , author Zhao, X. , author Dong, Z. , year 2023 . title Whu-helmet: A helmet-based multisensor slam dataset for the evaluation of real-time 3-d mapping in large-scale gnss-denied environments . journal IEEE Transactions on Geoscience and Remote Sensing volume 61 , pages 1--16

2023

[4] [4]

, author Huang, H

author Chen, X. , author Huang, H. , author Liu, Y. , author Li, J. , author Liu, M. , year 2022 . title Robot for automatic waste sorting on construction sites . journal Automation in Construction volume 141 , pages 104387

2022

[5] [5]

, author Yuan, S

author Li, J. , author Yuan, S. , author Cao, M. , author Nguyen, T.M. , author Cao, K. , author Xie, L. , year 2024 . title Hcto: Optimality-aware lidar inertial odometry with hybrid continuous time optimization for compact wearable mapping system . journal ISPRS Journal of Photogrammetry and Remote Sensing volume 211 , pages 228--243

2024

[6] [6]

, author Yang, B

author Li, J. , author Yang, B. , author Cong, Y. , author Cao, L. , author Fu, X. , author Dong, Z. , year 2019 . title 3d forest mapping using a low-cost uav laser scanning system: Investigation and comparison . journal Remote Sensing volume 11 , pages 717

2019

[7] [7]

, author Ramezani, M

author Proudman, A. , author Ramezani, M. , author Digumarti, S.T. , author Chebrolu, N. , author Fallon, M. , year 2022 . title Towards real-time forest inventory using handheld lidar . journal Robotics and Autonomous Systems volume 157 , pages 104240

2022

[8] [8]

, author Wang, W

author Wang, G. , author Wang, W. , author Ding, P. , author Liu, Y. , author Wang, H. , author Fan, Z. , author Bai, H. , author Hongbiao, Z. , author Du, Z. , year 2023 . title Development of a search and rescue robot system for the underground building environment . journal Journal of field robotics volume 40 , pages 655--683

2023