Decision Transformer-Based Drone Trajectory Planning with Dynamic Safety-Efficiency Trade-Offs

Chang-Hun Ji; SiWoon Song; SungTae Moon; Youn-Hee Han

REVIEW 4 major objections 6 minor 26 references

Decision Transformer-Based Drone Trajectory Planning with Dynamic Safety-Efficiency Trade-Offs

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A single Decision Transformer model can trade drone safety for speed by tuning one number.

desk verdict A useful, well-engineered application of Decision Transformers to drone planning, but the evidence for the claim that RTG is an intuitive safety-efficiency dial is thinner than the text suggests. read the letter →

arxiv 2507.21506 v2 pith:CEA7MXGW submitted 2025-07-29 cs.RO cs.AI

Chang-Hun Ji , SiWoon Song , Youn-Hee Han , SungTae Moon This is my paper · ORCID

classification cs.ROcs.AI

keywords DecisionTransformerReturn-to-Goconditioningsafety-efficiencytrade-offtrajectoryplanningquadrotornavigationofflinereinforcementlearningmedialaxisplannerN-stepreturn

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

The paper tries to establish that one trained Decision Transformer model can act as a drone trajectory planner whose safety–efficiency balance is controlled by a single scalar, the Return-to-Go (RTG). Its reward counts how many nearby obstacle voxels surround the drone, and because the training data come from expert planners that all reach the target, the paper reads high obstacle proximity as efficiency and low proximity as safety. Tuning RTG up or down then shifts the same model between obstacle-hugging fast routes and conservative clear routes, with no retraining and no reward redesign. Simulation on grid and random maps, plus real-world flights, is offered as evidence that the planner beats polynomial and reinforcement-learning baselines when tuned for either priority.

What carries the argument

The mechanism is a Decision Transformer conditioned on an N-step Return-to-Go, $G_t^N = \sum_{k=0}^{N-1} r_{t+k}$, with per-step reward $r_t = \sum_{i,j,k} o^{map}_t(i,j,k)\,\mathbf{1}(d_{ijk} < \delta)$ counting occupied voxels within radius $\delta$. The N-step truncation keeps the target return predictable in unknown environments, where the episode-long RTG of a standard Decision Transformer is ambiguous. Actions are the high-order coefficients of a fifth-order polynomial per axis, so the transformer only shapes the local trajectory while an autoencoder compresses the occupancy grid before conditioning. The trained model maps a desired return to the corresponding behavior mode, which is what lets a single weight set interpolate between safety and efficiency.

What would settle it

On the authors' expert dataset, compute the correlation between the per-step reward from Eq. (3) and flight time or average velocity; if high-reward segments are not systematically faster or shorter, the efficiency semantics of RTG collapse. A second check: in a single fixed map, sweep RTG and record clearance and flight time; if the two metrics do not move monotonically in opposite directions, the trade-off claim is not supported.

Watch

Extended reading notes

Core claim

The central discovery is that RTG can serve as a temperature parameter for trajectory style: a low RTG produces trajectories with larger clearance from obstacles, a high RTG produces faster, more direct trajectories, and the transition is available mid-flight by changing one input. This works because the Decision Transformer is trained offline on expert trajectories labeled by an N-step return that sums the number of obstacle voxels within a radius $\delta$; the paper argues that expert trajectories with many nearby obstacles are efficient because they are still on the way to the target. In the reported experiments, the low-RTG setting approaches the safety of a deliberate medial-axis reference planner, while the high-RTG setting achieves lower flight times than the polynomial baselines, and the ablation shows the N-step horizon is responsible for stable conditioning.

Load-bearing premise

That a trajectory with many nearby obstacles is genuinely more efficient, because every training trajectory comes from an expert planner that is heading to the target; the paper never measures the correlation between obstacle proximity and flight time or path length in its dataset.

Editorial extensions

If this is right

A user can switch mission priority mid-flight by changing the RTG input, without stopping the drone, retraining, or loading a second model.
RTG tuning needs no expert parameter knowledge because the same reward definition carries the same meaning across environments by counting obstacle proximity directly.
The N-step horizon is a functional component, not a detail: without it the conditioned trajectories become unstable or collide, and a too-long horizon blurs the trade-off.
The single model generalizes to unseen maps and to real hardware, so offline-expert data from multiple planners is enough to cover the behavior range.

Reading between the lines

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

The same conditioning idea could extend to other continuous trade-offs such as energy use, noise, or risk, by replacing the reward definition while keeping the RTG mechanism; the paper does not test this.
If the expert dataset contained suboptimal trajectories that hug obstacles without being fast, RTG would lose its efficiency meaning; this could be probed by adding noisy trajectories and checking whether high-RTG conditioning still speeds up flight.
Because the target return is measured in obstacle counts, the magnitude of a 'safe' or 'efficient' RTG likely depends on obstacle density; an automatic RTG selection rule based on local map statistics would be a natural next step.
The paper's mid-flight RTG switch suggests the model's behavior modes are encoded in the conditioning stream rather than learned as a hard policy, so trajectory style could be dialed by other auxiliary inputs as well.

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

Here's my take on the DT drone planner. The core idea—using RTG as a single dial for safety vs efficiency in one trained model—is a good practical application of Decision Transformers. The paper's real contributions are the specific engineering choices: the fifth-order polynomial action space, the N-step RTG truncation, and the medial-axis-based data augmentation for safe trajectories. Those are worth reporting, and the real-world demo strengthens the believability.

The central claim mostly holds up. In both grid and random maps, higher RTG produces faster flight and lower RTG produces safer distances, and evaluation metrics are separate from the reward, so there's no circularity. But the paper asks us to accept a relationship it never measures: that the Eq. (3) obstacle-count reward actually correlates with efficiency. The authors argue that expert trajectories are by definition efficient, so near-obstacle flight implies directness. That's a plausible heuristic, but it's an unverified one. A dataset-level analysis showing that N-step reward correlates with flight time or path length would close the gap. Without that, the claim that RTG is an intuitive, expert-free temperature is only an assertion.

The other soft spots are quantitative. Tables III and IV report averages of five runs with no error bars or significance tests. The differences between RTG=0.5 and RTG=10.0 are directionally consistent, but I can't tell if they're robust. The N-step ablation is even weaker: it's a visual comparison on a single grid map, and N=2 is selected from that one map. That's a shaky basis for a key design decision. And there's no code or data release, which limits reproducibility.

None of these are deal-breakers. They're all fixable, and the paper is a legitimate contribution to applied offline RL for robotics. I'd send it to review, but I'd ask for the variance statistics and the reward-efficiency correlation before accepting. The right audience is practitioners who want a single-model trade-off dial, and researchers working on RTG-conditioned policies.

Referee Report

4 major / 6 minor

Summary. The paper proposes a Decision Transformer (DT)-based drone trajectory planner in which a single Return-to-Go (RTG) parameter acts as a temperature controlling the safety–efficiency trade-off. The method represents local trajectories as fifth-order polynomial coefficient vectors, observes an occupancy grid and a local target vector, and is trained offline on expert data from EGO-Planner, ViGO, and a custom medial-axis-based planner. A reward function counts occupied voxels within a radius δ (Eq. 3), and an N-step return (Eq. 4) is used to make RTG conditioning practical in unknown environments. The authors report Gazebo simulations in grid and random maps showing that lower RTG gives safer trajectories and higher RTG gives more efficient ones, comparisons against EGO-Planner, ViGO, Safe EGO, and the medial-axis planner, an N-step RTG ablation, and a real-world flight demonstration.

Significance. If the central claim holds, the contribution is practically valuable: a user could change mission priorities during flight by tuning one interpretable parameter without retraining or redesigning a reward. The paper also adds a useful data-collection component by combining three expert planners, including a medial-axis planner for safe references, and it validates the idea in real flight. The evaluation is, however, currently built on an unmeasured assumed link between the Eq. (3) obstacle-count reward and flight efficiency, and the quantitative comparisons rest on five-run averages without variance or significance testing. The N-step RTG mechanism, which is central to the practical usability of the method, is justified by a single qualitative ablation. These issues are fixable with additional experiments and dataset-level analysis, so the paper merits a major revision rather than rejection.

major comments (4)

[§III.A.3, Eq. (3)] The reward function is defined solely as the number of occupied voxels within radius δ=20 of the drone. It does not include distance to the goal, path length, velocity, or flight time. The paper asserts in §III.A.3 that, because training data come from successful expert planners, many nearby obstacles indicate 'direct and efficient' flight while few obstacles indicate 'broader, safer' routes. This correlation is never measured on the collected dataset. A trajectory that meanders through dense clutter could accumulate high Eq. (3) reward while being slower than a clean, direct path through sparse regions. Since the central claim is that RTG tunes a safety–efficiency trade-off, please add dataset-level evidence: for the expert trajectories, report the correlation between per-trajectory (or per-N-step-window) Eq. (3) return and flight time and/or path length, or revise the reward to explicitly include a progress or time term. Without this, high-RTG conditioning may control obstacle proximity but not the claimed efficiency axis.
[§IV.B.4, Fig. 8] The N-step RTG ablation is performed on a single grid map (Grid map 4) and is qualitative: no flight time, safety metric, success rate, or quantitative trajectory comparison is reported for the w/o N-step, N=10, N=5, and N=2 conditions. The choice N=2 is then selected from this single-map comparison. Because the N-step return is a central component of the method, please evaluate N across multiple grid and random maps with quantitative metrics and report mean and variance. At minimum, report success/collision rates and the safety/efficiency metrics from Tables III and IV for each N.
[Tables III and IV, §IV.B.1–2] The tables report averages from five experiments without standard deviations, confidence intervals, or significance tests. Several headline comparisons are based on small point differences (e.g., Table III: Ours RTG=10.0 vs Safe EGO in flight time, 37.32 s vs 43.09 s; Table IV: Ours RTG=10.0 vs EGO in flight time, 39.47 s vs 42.13 s). Without variance or a statistical test, the claim that the proposed planner 'outperforms existing baseline methods' is not supported. Please add per-map results, standard deviations, and significance tests (or equivalent) for both safety and efficiency metrics.
[§IV.C, Fig. 10] The real-world experiment is descriptive: it shows stacked top-view images and velocity profiles but reports no quantitative safety metric (e.g., minimum distance to obstacles or obstacle count within a radius), no repeated trials, and no baseline comparison. Since the paper claims 'reliability and practicality' from this experiment, please add quantitative measures of safety and efficiency for the real-world runs, or soften the claim accordingly.

minor comments (6)

[Abstract/Introduction] There is a typo in the abstract and introduction: 'as atemperature parameter' should be 'as a temperature parameter'.
[§IV.B.1, Tables III and IV] Please state explicitly that for both safety metrics (DTW error and Obs Count) lower values are better, and note in Table III that 'Safety (DTW Error)' is computed against the medial-axis planner while Table IV uses obstacle counts.
[§IV.B.1, Fig. 5] The caption says 'Grid map 4' and 'Grid map 5' while the text mentions obstacle spacings of 4 m and 5 m; please clarify whether these are map indices or spacings, and ensure the axis labels and RTG values are readable in the figure.
[§IV.B.4, Fig. 8] The ablation figure would benefit from quantitative annotations (e.g., flight time, flight distance, safety metric) for each N condition, since the current qualitative description is hard to verify.
[§II, Related Work] The related-work section is adequate but could explicitly cite more recent offline RL and conditional sequence-modeling works to position the N-step RTG variant relative to existing return-conditioning methods.
[§III.A.2, Eq. (1)] Please clarify the trajectory duration or time horizon over which the fifth-order polynomial action is executed; this affects the interpretation of the replanning frequency and the N-step horizon in Eq. (4).

Circularity Check

1 steps flagged · score 2.0 of 10

One definitional conflation of obstacle proximity with efficiency; the empirical trade-off claim itself remains independently evaluated.

self definitional [Sec. III-B, 'Adjusting the Safety–Efficiency Trade-off via RTG' (following Sec. III-A.3, Eq. (3))]
"since our reward function assigns higher rewards to trajectories closer to obstacles (efficient trajectories) and lower rewards to trajectories farther from obstacles (safe trajectories), at inference time, a higher RTG instructs DT to select actions that prioritize efficiency."

The reward in Eq. (3) is defined solely as the number of occupied voxels within a threshold distance; it contains no flight-time, path-length, velocity, or distance-to-goal term. The paper's parenthetical labels obstacle-near trajectories as 'efficient trajectories,' so within the conditioning mechanism the efficiency effect of RTG is partly definitional: a higher RTG requests higher obstacle-count reward, and that obstacle count is renamed efficiency. The empirical evaluation uses separate metrics (flight time, average velocity, distance) that could falsify the proxy, so the central claim is not fully forced; however, the asserted link between Eq. (3) and real efficiency is an unmeasured assumption rather than a derived result.

full rationale

The paper does not claim a first-principles derivation of the safety–efficiency trade-off; it empirically trains a Decision Transformer on expert trajectories and conditions behavior on Return-to-Go. The central empirical claim is checked against external metrics: DTW error to the medial-axis planner in grid maps, obstacle count in random maps, and flight distance, average velocity, and flight time for efficiency. These metrics are not identical to the training reward, so the ordering in Tables III and IV is an experimental outcome rather than a mathematical identity. The only circularity-adjacent element is terminological: Sec. III-A.3 asserts that a high obstacle count indicates an efficient trajectory because the training data come from expert planners, and Sec. III-B then calls obstacle-near trajectories 'efficient trajectories' by definition. This makes the RTG-to-efficiency link partially self-definitional in the paper's own vocabulary, but because the evaluation uses independent efficiency metrics, the whole contribution does not reduce to its own definitions. There are no load-bearing self-citations or imported uniqueness theorems. The absence of variance or significance testing and the unvalidated proxy correlation are correctness risks, not circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The paper's trade-off control depends on hand-picked constants (delta, N, data-collection speed and frequency) and on several untested domain assumptions about what the reward proxy means. No new physical entities are introduced; N-step RTG is a modification of an existing construct.

free parameters (4)

Reward threshold delta = 20 (voxels)
Hand-chosen threshold in Eq. (3) defining the neighborhood around the drone for obstacle counting. It determines what 'near obstacle' means and shapes the reward signal; no sensitivity analysis is provided.
N-step return horizon N = 2
Selected after an ablation on Grid map 4 (N=10, N=5, N=2). It is a model hyperparameter controlling the RTG prediction horizon; no justification from theory or across environments is given.
Target velocity for data collection = 0.8 m/s
Constant used for all expert planners generating training data; affects the distribution of speed and efficiency and is not varied or justified.
Replanning frequency = 2 Hz
Rate at which expert planners generate new polynomial trajectories; affects temporal discretization, reward structure, and the meaning of an N-step return.

assumptions (5)

domain assumption Expert trajectories in the training dataset are collision-free and navigate toward the target, so high obstacle proximity correlates with efficient trajectories.
Used in Section III.A.3 to justify Eq. (3) as a safety-efficiency reward. If this correlation fails, high-RTG conditioning may produce unsafe or inefficient paths.
domain assumption A single DT model can produce diverse behaviors from different RTG conditioning values in unseen environments.
Core to the trade-off claim; relies on DT generalization properties and is validated only empirically in the paper.
domain assumption The medial-axis trajectory is the safest reference in grid maps.
Used to define DTW error as the safety metric in grid-map experiments; reasonable for structured layouts but not proven.
standard math The fifth-order polynomial action representation produces dynamically feasible local trajectories.
Assumed in Section III.A.2; standard for quadrotor trajectory planning, but the paper does not explicitly check feasibility constraints.
domain assumption The local occupancy grid and target vector form a sufficient state for planning.
The MDP observation in Section III.A.1 omits global map and history beyond the transformer context; assumes local perception suffices.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Decision Transformer-Based Drone Trajectory Planning with Dynamic Safety-Efficiency Trade-Offs." pith.science (2026). https://pith.science/paper/CEA7MXGW

@misc{pith2026250721506,
  author       = {Pith},
  title        = {Pith review of: Decision Transformer-Based Drone Trajectory Planning with Dynamic Safety-Efficiency Trade-Offs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CEA7MXGW}},
  note         = {Machine review of arXiv:2507.21506}
}

read the original abstract

A drone trajectory planner should be able to dynamically adjust the safety-efficiency trade-off according to varying mission requirements in unknown environments. Although traditional polynomial-based planners offer computational efficiency and smooth trajectory generation, they require expert knowledge to tune multiple parameters to adjust this trade-off. Moreover, even with careful tuning, the resulting adjustment may fail to achieve the desired trade-off. Similarly, although reinforcement learning-based planners are adaptable in unknown environments, they do not explicitly address the safety-efficiency trade-off. To overcome this limitation, we introduce a Decision Transformer-based trajectory planner that leverages a single parameter, Return-to-Go (RTG), as a \emph{temperature parameter} to dynamically adjust the safety-efficiency trade-off. In our framework, since RTG intuitively measures the safety and efficiency of a trajectory, RTG tuning does not require expert knowledge. We validate our approach using Gazebo simulations in both structured grid and unstructured random environments. The experimental results demonstrate that our planner can dynamically adjust the safety-efficiency trade-off by simply tuning the RTG parameter. Furthermore, our planner outperforms existing baseline methods across various RTG settings, generating safer trajectories when tuned for safety and more efficient trajectories when tuned for efficiency. Real-world experiments further confirm the reliability and practicality of our proposed planner.

Figures

Figures reproduced from arXiv: 2507.21506 by the authors.

**Figure 2.** An illustration of reward computation based on [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 4.** Examples of Gazebo simulation environments used [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗

Figures from the paper (5 more)

**Figure 5.** Figure 5: Comparison of trajectory generation and DTW error [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

**Figure 7.** Figure 7: Trajectory under RTG change in real-time. [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]

**Figure 8.** Figure 8: Trajectories with varying N in N-step RTG. efficiency metric is applied exactly as in the grid map environment. The random maps, which include environments unseen during training, were evaluated over five experiments, and the average performance is summarized in Tab. I…

**Figure 9.** Figure 9: The drone used in real-world experiments. [PITH_FULL_IMAGE:figures/full_fig_p007_9.png]

**Figure 10.** Figure 10: Results of real-world experiments. conditions: without N-step RTG (w/o N-step RTG), N = 10, N = 5, and our chosen experimental setting, N = 2. As shown in [PITH_FULL_IMAGE:figures/full_fig_p007_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 23 canonical work pages

[1]

Cognitive strategies for uav trajectory optimization: Ensuring safety and energy efficiency in real-world scenarios,

M. M. Umer, H. Venter, O. Muhammad, T. Shafique, F. A. Awwad, and E. A. Ismail, “Cognitive strategies for uav trajectory optimization: Ensuring safety and energy efficiency in real-world scenarios,” Ain Shams Engineering Journal , vol. 16, no. 3, p. 103301, 2025

work page 2025
[2]

Quadrotor path planning and polynomial trajectory generation using quadratic programming for indoor environments,

M. A. Arshad, J. Ahmed, and H. Bang, “Quadrotor path planning and polynomial trajectory generation using quadratic programming for indoor environments,” Drones, vol. 7, no. 2, p. 122, 2023

work page 2023
[3]

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

work page 2019
[4]

Learning minimum-time flight in cluttered environments,

R. Penicka, Y . Song, E. Kaufmann, and D. Scaramuzza, “Learning minimum-time flight in cluttered environments,” IEEE Robotics and Automation Letters, vol. 7, no. 3, pp. 7209–7216, 2022

work page 2022
[5]

Pathrl: An end-to-end path generation method for collision avoidance via deep reinforcement learning,

W. Yu, J. Peng, Q. Qiu, H. Wang, L. Zhang, and J. Ji, “Pathrl: An end-to-end path generation method for collision avoidance via deep reinforcement learning,” in 2024 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2024, pp. 9278–9284

work page 2024
[6]

Enhanc- ing safety via deep reinforcement learning in trajectory planning for agile flights in unknown environments,

L. Rocha, J. Bidinotto, F. Heintz, M. Tiger, and K. Vivaldini, “Enhanc- ing safety via deep reinforcement learning in trajectory planning for agile flights in unknown environments,” in 2024 IEEE/RSJ Interna- tional Conference on Intelligent Robots and Systems (IROS) . IEEE, 2024, pp. 3076–3083

work page 2024
[7]

Dynamic potential-based reward shaping,

S. M. Devlin and D. Kudenko, “Dynamic potential-based reward shaping,” in 11th International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2012) . IFAAMAS, 2012, pp. 433–440

work page 2012
[8]

Decision transformer: Reinforcement learning via sequence modeling,

L. Chen, K. Lu, A. Rajeswaran, K. Lee, A. Grover, M. Laskin, P. Abbeel, A. Srinivas, and I. Mordatch, “Decision transformer: Reinforcement learning via sequence modeling,” Advances in neural information processing systems , vol. 34, pp. 15 084–15 097, 2021

2021

Show all 26 references

[9]

Online decision transformer,

Q. Zheng, A. Zhang, and A. Grover, “Online decision transformer,” in international conference on machine learning . PMLR, 2022, pp. 27 042–27 059

2022
[10]

Elastic decision transformer,

Y .-H. Wu, X. Wang, and M. Hamaya, “Elastic decision transformer,” Advances in neural information processing systems , vol. 36, pp. 18 532–18 550, 2023

2023
[11]

Future-conditioned unsupervised pretraining for decision transformer,

Z. Xie, Z. Lin, D. Ye, Q. Fu, Y . Wei, and S. Li, “Future-conditioned unsupervised pretraining for decision transformer,” in International Conference on Machine Learning . PMLR, 2023, pp. 38 187–38 203

2023
[12]

Attention is all you need,

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” CoRR, vol. abs/1706.03762, 2017. [Online]. Available: http://arxiv.org/abs/1706.03762

2017 arXiv
[13]

Online trajectory planning for multiple quadrotors in dynamic environments using relative safe flight corridor,

J. Park and H. J. Kim, “Online trajectory planning for multiple quadrotors in dynamic environments using relative safe flight corridor,” IEEE Robotics and Automation Letters , vol. 6, no. 2, pp. 659–666, 2021

2021
[14]

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,” in 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . IEEE, 2022, pp. 6332– 6339

2022
[15]

Tgk-planner: An efficient topology guided kinodynamic planner for autonomous quadrotors,

H. Ye, X. Zhou, Z. Wang, C. Xu, J. Chu, and F. Gao, “Tgk-planner: An efficient topology guided kinodynamic planner for autonomous quadrotors,” IEEE Robotics and Automation Letters , vol. 6, no. 2, pp. 494–501, 2020

2020
[16]

Swarm of Micro Flying Robots in The Wild,

X. Zhou, X. Wen, Z. Wang, Y . Gao, H. Li, Q. Wang, T. Yang, H. Lu, Y . Cao, C. Xu, et al., “Swarm of Micro Flying Robots in The Wild,” Science Robotics, vol. 7, no. 66, p. eabm5954, 2022

2022
[17]

Vision-aided uav navigation and dynamic obstacle avoidance using gradient-based b- spline trajectory optimization,

Z. Xu, Y . Xiu, X. Zhan, B. Chen, and K. Shimada, “Vision-aided uav navigation and dynamic obstacle avoidance using gradient-based b- spline trajectory optimization,” in 2023 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2023, pp. 1214–1220

2023
[18]

Raptor: Robust and perception- aware trajectory replanning for quadrotor fast flight,

B. Zhou, J. Pan, F. Gao, and S. Shen, “Raptor: Robust and perception- aware trajectory replanning for quadrotor fast flight,” IEEE Transac- tions on Robotics , vol. 37, no. 6, pp. 1992–2009, 2021

1992
[19]

Au- tonomous Drone Racing with Deep Reinforcement Learning,

Y . Song, M. Steinweg, E. Kaufmann, and D. Scaramuzza, “Au- tonomous Drone Racing with Deep Reinforcement Learning,” in 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2021, pp. 1205–1212

2021
[20]

Deep reinforcement learning for drone navigation using sensor data,

V . J. Hodge, R. Hawkins, and R. Alexander, “Deep reinforcement learning for drone navigation using sensor data,” Neural Computing and Applications, vol. 33, no. 6, pp. 2015–2033, 2021

2015
[21]

Reinforcement learning-based multi-strategy cuckoo search algorithm for 3d uav path planning,

X. Yu and W. Luo, “Reinforcement learning-based multi-strategy cuckoo search algorithm for 3d uav path planning,” Expert Systems with Applications, vol. 223, p. 119910, 2023

2023
[22]

Efficient reinforcement learning for autonomous driving with parameterized skills and priors,

L. Wang, J. Liu, H. Shao, W. Wang, R. Chen, Y . Liu, and S. L. Waslan- der, “Efficient reinforcement learning for autonomous driving with parameterized skills and priors,” arXiv preprint arXiv:2305.04412 , 2023

2023 arXiv
[23]

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
[24]

G2VD planner: Efficient motion planning with grid-based generalized voronoi dia- grams,

J. Wen, X. Zhang, Q. Bi, H. Liu, J. Yuan, and Y . Fang, “G2VD planner: Efficient motion planning with grid-based generalized voronoi dia- grams,” IEEE Transactions on Automation Science and Engineering , 2024

2024
[25]

D+: A risk aware platform agnostic heterogeneous path planner,

S. Karlsson, A. Koval, C. Kanellakis, and G. Nikolakopoulos, “D+: A risk aware platform agnostic heterogeneous path planner,” Expert systems with applications , vol. 215, p. 119408, 2023

2023
[26]

Motion similarity evaluation between human and a tri-co robot during real-time imitation with a trajectory dynamic time warping model,

L. Gong, B. Chen, W. Xu, C. Liu, X. Li, Z. Zhao, and L. Zhao, “Motion similarity evaluation between human and a tri-co robot during real-time imitation with a trajectory dynamic time warping model,” Sensors, vol. 22, no. 5, p. 1968, 2022

1968

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Cognitive strategies for uav trajectory optimization: Ensuring safety and energy efficiency in real-world scenarios,

M. M. Umer, H. Venter, O. Muhammad, T. Shafique, F. A. Awwad, and E. A. Ismail, “Cognitive strategies for uav trajectory optimization: Ensuring safety and energy efficiency in real-world scenarios,” Ain Shams Engineering Journal , vol. 16, no. 3, p. 103301, 2025

work page 2025

[2] [2]

Quadrotor path planning and polynomial trajectory generation using quadratic programming for indoor environments,

M. A. Arshad, J. Ahmed, and H. Bang, “Quadrotor path planning and polynomial trajectory generation using quadratic programming for indoor environments,” Drones, vol. 7, no. 2, p. 122, 2023

work page 2023

[3] [3]

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

work page 2019

[4] [4]

Learning minimum-time flight in cluttered environments,

R. Penicka, Y . Song, E. Kaufmann, and D. Scaramuzza, “Learning minimum-time flight in cluttered environments,” IEEE Robotics and Automation Letters, vol. 7, no. 3, pp. 7209–7216, 2022

work page 2022

[5] [5]

Pathrl: An end-to-end path generation method for collision avoidance via deep reinforcement learning,

W. Yu, J. Peng, Q. Qiu, H. Wang, L. Zhang, and J. Ji, “Pathrl: An end-to-end path generation method for collision avoidance via deep reinforcement learning,” in 2024 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2024, pp. 9278–9284

work page 2024

[6] [6]

Enhanc- ing safety via deep reinforcement learning in trajectory planning for agile flights in unknown environments,

L. Rocha, J. Bidinotto, F. Heintz, M. Tiger, and K. Vivaldini, “Enhanc- ing safety via deep reinforcement learning in trajectory planning for agile flights in unknown environments,” in 2024 IEEE/RSJ Interna- tional Conference on Intelligent Robots and Systems (IROS) . IEEE, 2024, pp. 3076–3083

work page 2024

[7] [7]

Dynamic potential-based reward shaping,

S. M. Devlin and D. Kudenko, “Dynamic potential-based reward shaping,” in 11th International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2012) . IFAAMAS, 2012, pp. 433–440

work page 2012

[8] [8]

Decision transformer: Reinforcement learning via sequence modeling,

L. Chen, K. Lu, A. Rajeswaran, K. Lee, A. Grover, M. Laskin, P. Abbeel, A. Srinivas, and I. Mordatch, “Decision transformer: Reinforcement learning via sequence modeling,” Advances in neural information processing systems , vol. 34, pp. 15 084–15 097, 2021

2021