Pith. sign in

REVIEW 5 major objections 7 minor 36 references

Autonomous Vehicle Lateral Control Using Deep Reinforcement Learning with MPC-PID Demonstration

T0 review · 5 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Lateral control: a DRL agent can surpass the model-based controller that trained it.

desk verdict A plausible DRL-with-demonstration framework, but Algorithm 1 as written corrupts the replay buffer if implemented literally, and the experimental evidence is too thin to support the central claim. read the letter →

arxiv 2506.04040 v1 pith:DLEH3SFN submitted 2025-06-04 cs.RO cs.LGcs.SYeess.SY

Chengdong Wu , Sven Kirchner , Nils Purschke , Alois C. Knoll This is my paper
classification cs.ROcs.LGcs.SYeess.SY
keywords autonomousdrivinglateralcontroldeepreinforcementlearningDDPGmodelpredictivePIDcontrollerfromdemonstrationpathfollowing
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 claims that a deep reinforcement learning (DRL) controller can outperform the model-based controller that served as its trainer, provided the demonstration is fed in online during training. Specifically, a Deep Deterministic Policy Gradient (DDPG) agent is trained for vehicle lateral control while an MPC-PID controller runs in parallel, contributing both a reward term that penalizes deviation from the demonstrator's steering angle and a probabilistic extra action. If the claim holds, suboptimal model-based controllers can act as training guides for model-free policies that then surpass them, reducing manual tuning and improving robustness to incomplete vehicle information. The claims are supported by experiments in a driving simulator on circular and racing tracks, including tests with corrupted waypoints.

What carries the argument

The load-bearing mechanism is the online demonstration loop: at every step the MPC-PID controller outputs a steering suggestion, and this suggestion enters the DDPG update in two ways. First, the reward function adds a penalty term $-\frac{1}{2} c_{\mathrm{diff}} (\delta_{\mathrm{demo}} - \delta)^2$ to the path-tracking reward, with time-varying coefficients $c_{\mathrm{track}}$ and $c_{\mathrm{diff}}$ set by sigmoid functions of the corresponding reward components, so that the agent trusts the demonstrator early and shifts to environment reward later. Second, with probability $p_{\mathrm{action}}$ the demonstrator's action is executed in the environment while the policy's sampled action is stored in the replay buffer. A cyclical learning-rate schedule is used to speed and stabilize convergence.

What would settle it

Re-run the training with the pseudocode's logic instrumented: if the reward and executed action are mismatched on demonstration steps, then a version that stores the actually executed action should produce degraded or unstable learning; conversely, if both versions learn equally well, the mismatch is irrelevant. A simpler check is to inspect the replay buffer during training and confirm that stored actions match executed actions on demonstration steps.

Watch

Extended reading notes

Core claim

The central discovery is that a DDPG policy trained with an online MPC-PID demonstrator achieves lower path-following error than the demonstrator itself, especially when waypoint information is corrupted by noise, and it does so with smoother, faster convergence than plain DDPG. The agent observes ground-truth waypoints and vehicle states, and its steering output combines the learned policy with the demonstrator's suggestion during training. After training, the demonstrator is detached, and the learned policy alone guides the vehicle. Reported results include an average lateral error below 0.104 meters on a circular track at full throttle, and on an unfamiliar racing track the learned controller uses about 22% of the computational time of the MPC-PID controller while completing the lap.

Load-bearing premise

The training result depends on every stored transition pairing the executed action with the reward it produced; Algorithm 1 appears to store the sampled policy action even when the demonstrator's action was executed, so the claimed training benefit may rest on an incorrect replay buffer.

Editorial extensions

If this is right

  • If the claim holds, any reasonably competent model-based controller can be used as an online demonstrator to bootstrap a model-free policy that outperforms it, making the approach a general recipe for upgrading suboptimal controllers.
  • The trained controller runs without the demonstrator, so control cost is a single neural-network evaluation rather than online optimization, with a reported 78% reduction in computational time per lap.
  • Robustness to corrupted waypoints suggests the controller tolerates perception or planning errors in the driving pipeline, simplifying integration.
  • The adaptive weighting between imitation and environment reward yields smooth convergence, indicating that relative trust in a demonstrator can be learned rather than hand-tuned.

Reading between the lines

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

  • A direct test of Algorithm 1's replay-buffer logic would settle whether the reported benefit survives a corrected implementation: if the executed action (demonstrator or policy) is stored instead of the sampled policy action, convergence may degrade, which would indicate the gains came partly from an off-policy bias.
  • The online-demonstration recipe likely transfers to other continuous-control tasks with a reliable but suboptimal base controller, such as trajectory tracking for manipulators or drones, though the paper only demonstrates lateral vehicle control.
  • Comparing against a pretraining-only baseline (training exclusively on collected demonstrations before environment interaction) would quantify the specific advantage of online, synchronous demonstration over offline imitation pretraining.
  • The time-varying reward coefficients could be made adaptive to the demonstrator's measured performance rather than to reward magnitudes, which might improve robustness when the demonstrator is unreliable in some regimes.
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.

Referee Report

5 major / 7 minor

Summary. The manuscript proposes a Deep Deterministic Policy Gradient (DDPG) lateral controller for autonomous vehicles trained in CARLA with an MPC-PID controller acting as an online demonstrator. The demonstrator contributes a revised reward term penalizing deviation from the demonstrator's steering action and, with probability p_action, directly supplies the executed steering angle. After training, the MPC-PID controller is detached and the DRL agent is tested on a circular track and a Monza-shaped racing track, with and without waypoint noise. The paper claims that the DRL agent surpasses the MPC-PID demonstrator in path-following accuracy, especially under incomplete vehicle information, and that the online demonstration stabilizes and accelerates DDPG training.

Significance. The proposed integration of model-based control as an online demonstration source is practically motivated and could lower the exploration burden in DRL-based lateral control if the training signal is correctly formed. The experimental setup—CARLA, multiple track shapes, and waypoint-noise robustness tests—is appropriate for a first feasibility study, and the detach-after-training scheme is clearly described. However, the current evidence is not sufficient to support the strong comparative claims: the formal training algorithm contains a potentially invalid off-policy update, the MPC-PID baseline is not evaluated quantitatively on the same metrics, and all results appear to be single-run with no error bars or released code.

major comments (5)
  1. [III.C, Algorithm 1] In Algorithm 1, after sampling the policy action a, the agent executes the demonstrator action a_demo with probability p_action, but the transition stored in the replay buffer is (s, a, r, s', d) with the sampled policy action rather than the executed action. When a_demo is executed, the reward r and next state s' are produced by a_demo; the subsequent Q-function update therefore trains Q(s,a) on a target that is a function of a different action's outcome. This violates the off-policy correctness condition of DDPG and, if the implementation follows the pseudocode literally, the smooth actor-loss curves in Fig. 4 cannot be interpreted as evidence for the proposed demonstration mechanism. Please either correct the pseudocode to store the executed action, or, if the implementation already does so, state this explicitly and release the code so the discrepancy can be resolved.
  2. [V.A, Table II and Fig. 3] The central claim that the DRL controller 'surpasses the performance of the demonstrating controller' is not supported by quantitative evidence. Table II reports ALE/AOE only for the DRL agent on the circular track, and the racing-track comparison in Fig. 3 is qualitative. The only quantitative information about MPC-PID is the wall-clock comparison (145.28 s vs. 664.23 s), which does not measure path-following error. Please provide ALE/AOE, orientation error, and track-completion results for the MPC-PID demonstrator on the same tracks, speeds, and waypoint-noise conditions as the DRL agent.
  3. [V.B, Fig. 4] All training curves and test numbers appear to come from a single run. DDPG with OU noise and cyclical learning rates is stochastic, so the claimed 'smooth descending profile' and training stabilization cannot be distinguished from run-to-run variation without multiple seeds. Please report mean ± standard deviation over at least three independent training runs for the actor loss, and also report the variance of the evaluation metrics in Table II.
  4. [III.C, Eqs. (6)-(8); V.B, Fig. 5] The adaptive reward-weighting scheme is internally inconsistent as presented. If c_track and c_diff are normalized sigmoids of r_track and r_diff as in Eqs. (7)-(8), then at the start of training, when the random policy typically produces large lateral error and large deviation from the demonstrator, r_track is small or negative while r_diff is large, so c_diff should dominate. The text in Section V.B states that 'the track reward possesses a larger proportion during the initial 100,000 training steps' but then concludes that this 'indicates more trust in imitation learning... at the beginning.' These statements contradict each other and the equations. Please clarify the normalization, report the actual coefficient profiles, and align the narrative with Fig. 5.
  5. [III.C-IV] The experiments are not reproducible from the text: no values are given for the DDPG network sizes, batch size, γ, τ, OU noise parameters, p_action, the reward coefficients before normalization, or the MPC weights Q, P, r, T_p, c_MPC, and c_PID in Eq. (2). These quantities are central to the claimed training behavior and to the MPC-PID baseline. Please include a complete hyperparameter table and, ideally, the code or a public repository.
minor comments (7)
  1. [II.C] The sentence 'The DDPG will be used as the baseline of the baseline of the RL algorithm applied in this paper' contains a duplicated phrase and should be corrected.
  2. [V.A] 'A different test rack' should be 'a different test track'; the paper also uses 'circuit' and 'track' interchangeably in the same paragraph.
  3. [Table II] The columns report throttle position, but the text discusses velocity; please clarify the relationship between throttle and speed and specify the actual vehicle speed for each row.
  4. [Fig. 2] The waypoint-noise robustness results are presented only as plots; provide ALE/AOE values for the noisy condition so the robustness claim can be quantified.
  5. [Eq. (1)] The state vector is defined as X = [y, β, ψ, ψdot], but the cost function uses (x, y) positions; please clarify the state-space definition and the relationship between lateral position y and the x/y coordinates.
  6. [III.C] The statement that 'because DDPG outputs a concrete control signal, the inherent exploration characteristics does not exist' is inaccurate; DDPG has a deterministic policy but relies on injected noise for exploration. Please rephrase.
  7. [Figs. 4 and 5] The axes and legends in Figs. 4 and 5 are not fully described; please add axis labels, units, and clarify the legend entries, and reconcile 'cdemo' in Fig. 5 with 'cdiff' in the text.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the DRL-vs-MPC-PID comparison is an empirical benchmark, and no derived quantity reduces to an input by construction.

full rationale

The paper's central claim is empirical: a DDPG controller trained with online MPC-PID demonstration is reported to outperform the demonstrating MPC-PID controller in CARLA. This claim is tested against the simulator and against conventional DRL training, not derived from a fitted parameter or from an equation that contains the conclusion. The revised reward function in Eq. (5) adds an imitation term, but it also retains the track-following reward and uses time-varying coefficients (Eqs. (7)-(8)), so the learning outcome is not forced by construction; the agent could in principle fail to match or exceed the demonstrator. There are no load-bearing self-citations: the cited works are external baselines, algorithms, and simulators. Comparing the trained agent with its own demonstrator is a benchmark-design choice rather than circular reasoning, because the agent is detached from the demonstrator during testing and must generalize to new tracks and noisy waypoints. The replay-buffer mismatch in Algorithm 1, where the sampled policy action is stored even when the MPC-PID action is executed, is a serious correctness risk for the training claim if the implementation follows the pseudocode literally, but it is not circularity: it does not make any prediction equivalent to its input by definition. The paper is therefore self-contained against the CARLA environment and external DDPG machinery, with no circular step to report.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests on several hand-chosen parameters and domain assumptions: the quality of the MPC-PID demonstrator, the treatment of ground-truth waypoints plus noise as realistic perception error, the ad hoc sigmoid reward-weighting scheme, and the validity of CARLA as an evaluation environment. No new physical entities, forces, or conserved quantities are introduced.

free parameters (6)
  • MPC cost weights Q, P, r and prediction horizon Tp
    Chosen by hand in Sec. III-B; the authors state the MPC and PID hyperparameters are tuned together but give no values or tuning procedure.
  • PID feedback gains and combination weights cMPC, cPID in Eq. (2)
    Tuned together with the MPC parameters; no values are reported.
  • Reward coefficients ctrack, cdiff, and cchange in Eqs. (4) and (5)
    Hand-designed weights; initial values and normalization details are not reported.
  • DDPG hyperparameters including network sizes, batch size, gamma, tau, Ornstein-Uhlenbeck noise, and cyclical learning…
    Required for reproduction; not reported beyond the hardware table and the use of Stable Baselines3.
  • Demonstration execution probability paction
    Predefined probability in Algorithm 1; its value is not given.
  • Number and spacing of lookahead waypoints observed by the agent
    Observation design choice; the paper only says a high number of waypoints without specifying the exact count.
assumptions (5)
  • domain assumption The MPC uses a time-invariant continuous linear vehicle model with states [y, beta, psi, psi_dot], and this simplified model is adequate for the baseline despite acknowledged inaccuracies.
    Invoked in Sec. III-B; the authors themselves state that the simplified model and limited parameter accuracy lead to unacceptable control performance, which motivates the PID feedback and DRL.
  • domain assumption Ground-truth CARLA waypoints are treated as the desired path, replacing detection and planning modules, and random waypoint noise is taken to represent perception or planning errors.
    Stated in Sec. III-C and Sec. IV; the robustness claim depends on this equivalence.
  • ad hoc to paper The sigmoid-normalized time-varying reward coefficients in Eqs. (6) to (8) appropriately balance imitation and tracking during training.
    No proof or ablation is provided; the paper only plots the coefficient profiles as evidence that the mechanism works as demanded.
  • domain assumption CARLA simulation with a fixed 0.05 s synchronous timestep and a 3 m corridor is a valid surrogate for evaluating lateral control performance.
    Used throughout Secs. IV and V; no real-vehicle validation is presented.
  • standard math DDPG with a replay buffer and target networks converges to a good policy in this MDP.
    Standard assumption for DDPG; not proven for this specific MDP and observation space.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Autonomous Vehicle Lateral Control Using Deep Reinforcement Learning with MPC-PID Demonstration." pith.science (2026). https://pith.science/paper/DLEH3SFN

@misc{pith2026250604040,
  author       = {Pith},
  title        = {Pith review of: Autonomous Vehicle Lateral Control Using Deep Reinforcement Learning with MPC-PID Demonstration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DLEH3SFN}},
  note         = {Machine review of arXiv:2506.04040}
}
read the original abstract

The controller is one of the most important modules in the autonomous driving pipeline, ensuring the vehicle reaches its desired position. In this work, a reinforcement learning based lateral control approach, despite the imperfections in the vehicle models due to measurement errors and simplifications, is presented. Our approach ensures comfortable, efficient, and robust control performance considering the interface between controlling and other modules. The controller consists of the conventional Model Predictive Control (MPC)-PID part as the basis and the demonstrator, and the Deep Reinforcement Learning (DRL) part which leverages the online information from the MPC-PID part. The controller's performance is evaluated in CARLA using the ground truth of the waypoints as inputs. Experimental results demonstrate the effectiveness of the controller when vehicle information is incomplete, and the training of DRL can be stabilized with the demonstration part. These findings highlight the potential to reduce development and integration efforts for autonomous driving pipelines in the future.

Figures

Figures reproduced from arXiv: 2506.04040 by the authors.

Figure 1
Figure 1. Overview of the proposed DRL controller with demonstration. The real steering command for the vehicle comes [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Test data for lateral error (left) and orientation error [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 5
Figure 5. ctrack (left) and cdemo (right) during the training process. VI. CONCLUSIONS We propose a reinforcement learning framework with an online demonstration from the MPC-PID controller to combine the advantages of imitation learning and reinforce￾ment learning. The MPC-PID demonstration is used for revising the reward function and taking extra action for exploration. A time-variant reward function is developed to adapt t… view at source ↗
Figures from the paper (2 more)
Figure 3
Figure 3. Figure 3: Relative position between ego-vehicle and the desired [PITH_FULL_IMAGE:figures/full_fig_p007_3.png]
Figure 4
Figure 4. Figure 4: Actor losses for different approaches with and without [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 32 canonical work pages

  1. [32]

    A Deep Reinforcement Learning Algorithm with Expert Demonstrations and Su- pervised Loss and its application in Autonomous Driving

    Kai Liu, Qin Wan, and Yanjie Li. “A Deep Reinforcement Learning Algorithm with Expert Demonstrations and Su- pervised Loss and its application in Autonomous Driving”. In: 2018 37th Chinese Control Conference (CCC) . 2018, pp. 2944–2949

    work page 2018

  2. [33]

    Improved Deep Reinforcement Learning with Expert Demonstrations for Urban Autonomous Driv- ing

    Haochen Liu et al. “Improved Deep Reinforcement Learning with Expert Demonstrations for Urban Autonomous Driv- ing”. In: 2022 IEEE Intelligent Vehicles Symposium (IV) . 2022, pp. 921–928

    work page 2022

  3. [1]

    SAE Standards News: J3016 automated- driving graphic update

    SAE International. SAE Standards News: J3016 automated- driving graphic update

    work page

  4. [2]

    The State-of-the-art of Model Predictive Control in Recent Years

    Jixia Han, Yi Hu, and Songyi Dian. “The State-of-the-art of Model Predictive Control in Recent Years”. In: IOP Conference Series: Materials Science and Engineering 428 (2018), p. 012035

    work page 2018

  5. [3]

    Au- tomated Tuning of Nonlinear Model Predictive Controller by Reinforcement Learning

    Mohit Mehndiratta, Efe Camci, and Erdal Kayacan. “Au- tomated Tuning of Nonlinear Model Predictive Controller by Reinforcement Learning”. In: 2018 IEEE/RSJ Interna- tional Conference on Intelligent Robots and Systems (IROS) (Madrid). IEEE, 2018, pp. 3016–3021. ISBN : 978-1-5386- 8094-0

    work page 2018

  6. [4]

    Reinforcement Learning of the Pre- diction Horizon in Model Predictive Control

    Eivind Bøhn et al. “Reinforcement Learning of the Pre- diction Horizon in Model Predictive Control”. In: IFAC- PapersOnLine 54.6 (2021), pp. 314–320. ISSN : 2405-8963

    work page 2021

  7. [5]

    Trajectory Tracking of Autonomous Vehicle Based on Model Predictive Control With PID Feed- back

    Duanfeng Chu et al. “Trajectory Tracking of Autonomous Vehicle Based on Model Predictive Control With PID Feed- back”. In: IEEE Transactions on Intelligent Transportation Systems 24.2 (2023), pp. 2239–2250

    work page 2023

  8. [6]

    Imitation Learning in Robots

    Billard, Aude and Grollman, Daniel. “Imitation Learning in Robots”. In: Encyclopedia of the Sciences of Learning . Ed. by Norbert M. Seel. Boston, MA: Springer US, 2012, pp. 1494–1496. ISBN : 978-1-4419-1428-6

    work page 2012

Show all 36 references
  1. [7]

    Sutton and Andrew Barto

    Richard S. Sutton and Andrew Barto. Reinforcement learn- ing: An introduction. Second edition. Adaptive computation and machine learning. Cambridge, Massachusetts and Lon- don, England: The MIT Press, 2020. ISBN : 9780262039246

    2020

  2. [8]

    CARLA: An Open Urban Driving Simulator

    Alexey Dosovitskiy et al. “CARLA: An Open Urban Driving Simulator”. In: CoRR abs/1711.03938 (2017)

    2017 arXiv

  3. [9]

    Lateral control for autonomous wheeled vehicles: A technical review

    Yassine Kebbati et al. “Lateral control for autonomous wheeled vehicles: A technical review”. In: Asian Journal of Control 25.4 (2023), pp. 2539–2563

    2023

  4. [10]

    Adaptive LQR Path Tracking Control for 4WS Electric Vehicles Based on Genetic Algorithm

    Ao Lu et al. “Adaptive LQR Path Tracking Control for 4WS Electric Vehicles Based on Genetic Algorithm”. In: 2022 6th CAA International Conference on Vehicular Control and Intelligence (CVCI). 2022, pp. 1–6

    2022

  5. [11]

    Development of a Sliding- Mode-Control-Based Path-Tracking Algorithm with Model- Free Adaptive Feedback Action for Autonomous Vehicles

    Kwangseok Oh and Jaho Seo. “Development of a Sliding- Mode-Control-Based Path-Tracking Algorithm with Model- Free Adaptive Feedback Action for Autonomous Vehicles”. In: Sensors 23.1 (2023). ISSN : 1424-8220

    2023

  6. [12]

    Extension of the law of large numbers to quantities, depending on each other (1906). Reprint

    A. A. Markov. “Extension of the law of large numbers to quantities, depending on each other (1906). Reprint”. In: Journal ´Electronique d’Histoire des Probabilit ´es et de la Statistique [electronic only] 2.1b (2006), Article 10, 12, electronic only–Article 10, 12, electronic only

    1906

  7. [13]

    Playing Atari with Deep Reinforce- ment Learning

    V olodymyr Mnih et al. Playing Atari with Deep Reinforce- ment Learning. 2013

    2013

  8. [14]

    Deterministic policy gradient algo- rithms

    David Silver et al. “Deterministic policy gradient algo- rithms”. In: Proceedings of the 31st International Confer- ence on International Conference on Machine Learning - Volume 32. ICML’14. JMLR.org, 2014, pp. I–387–I–395

    2014

  9. [15]

    Continuous control with deep reinforcement learning

    Timothy P. Lillicrap et al. “Continuous control with deep reinforcement learning”. In: CoRR abs/1509.02971 (2015)

    2015 arXiv

  10. [16]

    Proximal Policy Optimization Algo- rithms

    John Schulman et al. “Proximal Policy Optimization Algo- rithms”. In: CoRR abs/1707.06347 (2017)

    2017 arXiv

  11. [17]

    Soft actor-critic: Off-policy maxi- mum entropy deep reinforcement learning with a stochastic actor

    Tuomas Haarnoja et al. “Soft actor-critic: Off-policy maxi- mum entropy deep reinforcement learning with a stochastic actor”. In: International Conference on Machine Learning (ICML) (2018)

    2018

  12. [18]

    Reinforcement-Learning-Based Coop- erative Adaptive Cruise Control of Buses in the Lincoln Tunnel Corridor with Time-Varying Topology

    Weinan Gao et al. “Reinforcement-Learning-Based Coop- erative Adaptive Cruise Control of Buses in the Lincoln Tunnel Corridor with Time-Varying Topology”. In: IEEE Transactions on Intelligent Transportation Systems 20.10 (2019), pp. 3796–3805

    2019

  13. [19]

    Deep reinforcement learning based control for Autonomous Vehicles in CARLA

    ´Oscar P ´erez-Gil et al. “Deep reinforcement learning based control for Autonomous Vehicles in CARLA”. In: Multi- media Tools and Applications 81.3 (2022), pp. 3553–3576. ISSN : 1573-7721

    2022

  14. [20]

    SRL-TR2: A Safe Reinforcement Learning Based TRajectory TRacker Framework

    Chengyu Wang et al. “SRL-TR2: A Safe Reinforcement Learning Based TRajectory TRacker Framework”. In: IEEE Transactions on Intelligent Transportation Systems 24.6 (2023), pp. 5765–5780

    2023

  15. [21]

    Deep Reinforcement Learning Based Tracking Control of Un- manned Vehicle with Safety Guarantee

    Zhongjing Luo, Jialing Zhou, and Guanghui Wen. “Deep Reinforcement Learning Based Tracking Control of Un- manned Vehicle with Safety Guarantee”. In:2022 13th Asian Control Conference (ASCC). 2022, pp. 1893–1898

    2022

  16. [22]

    Autonomous Automobile Tra- jectory Tracking for Off-Road Driving: Controller Design, Experimental Validation and Racing

    Gabriel M. Hoffmann et al. “Autonomous Automobile Tra- jectory Tracking for Off-Road Driving: Controller Design, Experimental Validation and Racing”. In: 2007 American Control Conference. 2007, pp. 2296–2301

    2007

  17. [23]

    Path-Tracking Control Strategy of Unmanned Vehicle Based on DDPG Algorithm

    Jialing Yao and Zhen Ge. “Path-Tracking Control Strategy of Unmanned Vehicle Based on DDPG Algorithm”. In:Sensors 22.20 (2022). ISSN : 1424-8220

    2022

  18. [24]

    Self-Optimizing Path Tracking Controller for Intelligent Vehicles Based on Reinforcement Learning

    Jichang Ma et al. “Self-Optimizing Path Tracking Controller for Intelligent Vehicles Based on Reinforcement Learning”. In: Symmetry 14.1 (2022). ISSN : 2073-8994

    2022

  19. [25]

    Reinforcement Learning of Model Predic- tive Control Parameters for Autonomous Vehicle Guidance

    Baha Zarrouki. “Reinforcement Learning of Model Predic- tive Control Parameters for Autonomous Vehicle Guidance”. Unpublished, 2020

    2020

  20. [26]

    A Combined Reinforcement Learning and Model Predictive Control for Car-Following Maneuver of Autonomous Vehicles

    Liwen Wang et al. “A Combined Reinforcement Learning and Model Predictive Control for Car-Following Maneuver of Autonomous Vehicles”. In: Chinese Journal of Mechan- ical Engineering 36.1 (2023), p. 80. ISSN : 2192-8258

    2023

  21. [27]

    Model Predictive Control for Reli- able Path Following with Application to the Autonomous Vehicle and Considering Different Vehicle Models

    Behnaz Ahmadi et al. “Model Predictive Control for Reli- able Path Following with Application to the Autonomous Vehicle and Considering Different Vehicle Models”. In: 2021 5th International Conference on Vision, Image and Signal Processing (ICVISP) . 2021, pp. 27–32

    2021

  22. [28]

    Multi- ple Model Predictive Control: A State Estimation based Approach

    Matthew Kuure-Kinsey and B. Wayne Bequette. “Multi- ple Model Predictive Control: A State Estimation based Approach”. In: 2007 American Control Conference . 2007, pp. 3739–3744

    2007

  23. [29]

    Model Predictive Control With Learned Vehicle Dynamics for Autonomous Vehicle Path Tracking

    Mohammad Rokonuzzaman et al. “Model Predictive Control With Learned Vehicle Dynamics for Autonomous Vehicle Path Tracking”. In: IEEE Access 9 (2021), pp. 128233– 128249

    2021

  24. [30]

    Model- free reinforcement learning from expert demonstrations: a survey

    Jorge Ram ´ırez, Wen Yu, and Adolfo Perrusqu ´ıa. “Model- free reinforcement learning from expert demonstrations: a survey”. In: Artificial Intelligence Review 55.4 (2022), pp. 3213–3241. ISSN : 0269-2821

    2022

  25. [31]

    Integrating reinforcement learning with human demon- strations of varying ability

    Matthew E. Taylor, Halit Bener Suay, and Sonia Chernova. “Integrating reinforcement learning with human demon- strations of varying ability”. In: The 10th International Conference on Autonomous Agents and Multiagent Systems - Volume 2. AAMAS ’11. Richland, SC: International Fo...

    2011

  26. [34]

    Learning from Demonstrations for Real World Reinforcement Learning

    Todd Hester et al. “Learning from Demonstrations for Real World Reinforcement Learning”. In: CoRR abs/1704.03732 (2017)

    2017 arXiv

  27. [35]

    Stable-Baselines3: Reliable Reinforce- ment Learning Implementations

    Antonin Raffin et al. “Stable-Baselines3: Reliable Reinforce- ment Learning Implementations”. In: Journal of Machine Learning Research 22.268 (2021), pp. 1–8

    2021

  28. [36]

    Deep Reinforcement Learning using Cyclical Learning Rates

    Ralf Gulde et al. “Deep Reinforcement Learning using Cyclical Learning Rates”. In: CoRR abs/2008.01171 (2020)

    2020 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.