Skip to main content

Thank you for visiting You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

Continuous improvement of self-driving cars using dynamic confidence-aware reinforcement learning


Today’s self-driving vehicles have achieved impressive driving capabilities, but still suffer from uncertain performance in long-tail cases. Training a reinforcement-learning-based self-driving algorithm with more data does not always lead to better performance, which is a safety concern. Here we present a dynamic confidence-aware reinforcement learning (DCARL) technology for guaranteed continuous improvement. Continuously improving means that more training always improves or maintains its current performance. Our technique enables performance improvement using the data collected during driving, and does not need a lengthy pre-training phase. We evaluate the proposed technology both using simulations and on an experimental vehicle. The results show that the proposed DCARL method enables continuous improvement in various cases, and, in the meantime, matches or outperforms the default self-driving policy at any stage. This technology was demonstrated and evaluated on the vehicle at the 2022 Beijing Winter Olympic Games.

This is a preview of subscription content, access via your institution

Access options

Rent or buy this article

Get just this article for as long as you need it


Prices may be subject to local taxes which are calculated during checkout

Fig. 1: Continuous improvement concept and framework of the DCARL agent.
Fig. 2: Continuous improvement process of the DCARL agent.
Fig. 3: Continuously improved performance with more data.
Fig. 4: DCARL agent for long-tail cases.
Fig. 5: Experimental results.
Fig. 6: Data for evaluation and theoretical analysis.

Data availability

The Supplementary Software file contains the minimum data to run and render the results for all three experiments. These data are also available in a public repository at (ref. 37).

Code availability

The source code of the self-driving experiments is available at (ref. 38). It contains the proposed DCARL planning algorithms as well as the used perception, localization and control algorithms in our self-driving cars.


  1. Sutton, R. S. & Barto, A. G. Reinforcement Learning: An Introduction (MIT Press, 2018).

  2. Silver, D. et al. A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play. Science 362, 1140–1144 (2018).

    Article  MATH  MathSciNet  Google Scholar 

  3. Silver, D. et al. Mastering the game of Go with deep neural networks and tree search. Nature 529, 484–489 (2016).

    Article  Google Scholar 

  4. Mnih, V. et al. Human-level control through deep reinforcement learning. Nature 518, 529–533 (2015).

    Article  Google Scholar 

  5. Ye, F., Zhang, S., Wang, P. & Chan, C.-Y. A survey of deep reinforcement learning algorithms for motion planning and control of autonomous vehicles. In 2021 IEEE Intelligent Vehicles Symposium (IV) 1073–1080 (IEEE, 2021).

  6. Zhu, Z. & Zhao, H. A survey of deep RL and IL for autonomous driving policy learning. IEEE Trans. Intell. Transp. Syst. 23, 14043–14065 (2022).

  7. Aradi, S. Survey of deep reinforcement learning for motion planning of autonomous vehicles. IEEE Trans. Intell. Transp. Syst. 23, 740–759 (2022).

    Article  Google Scholar 

  8. Cao, Z. et al. Highway exiting planner for automated vehicles using reinforcement learning. IEEE Trans. Intell. Transp. Syst. 22, 990–1000 (2020).

    Article  Google Scholar 

  9. Stilgoe, J. Self-driving cars will take a while to get right. Nat. Mach. Intell. 1, 202–203 (2019).

    Article  Google Scholar 

  10. Kalra, N. & Paddock, S. M. Driving to safety: How many miles of driving would it take to demonstrate autonomous vehicle reliability? Transp. Res. Part A 94, 182–193 (2016).

    Google Scholar 

  11. Disengagement reports. California DMV (2021).

  12. Li, G. et al. Decision making of autonomous vehicles in lane change scenarios: deep reinforcement learning approaches with risk awareness. Transp. Res. Part C 134, 103452 (2022).

    Article  Google Scholar 

  13. Shu, H., Liu, T., Mu, X. & Cao, D. Driving tasks transfer using deep reinforcement learning for decision-making of autonomous vehicles in unsignalized intersection. IEEE Trans. Veh. Technol. 71, 41–52 (2021).

    Article  Google Scholar 

  14. Pek, C., Manzinger, S., Koschi, M. & Althoff, M. Using online verification to prevent autonomous vehicles from causing accidents. Nat. Mach. Intell. 2, 518–528 (2020).

    Article  Google Scholar 

  15. Xu, S., Peng, H., Lu, P., Zhu, M. & Tang, Y. Design and experiments of safeguard protected preview lane keeping control for autonomous vehicles. IEEE Access 8, 29944–29953 (2020).

    Article  Google Scholar 

  16. Yang, J., Zhang, J., Xi, M., Lei, Y. & Sun, Y. A deep reinforcement learning algorithm suitable for autonomous vehicles: double bootstrapped soft-actor-critic-discrete. IEEE Trans. Cogn. Dev. Syst. (2021).

  17. Schwall, M., Daniel, T., Victor, T., Favaro, F. & Hohnhold, H. Waymo public road safety performance data. Preprint at arXiv (2020).

  18. Fan, H. et al. Baidu Apollo EM motion planner. Preprint at arXiv (2018).

  19. Kato, S. et al. Autoware on board: enabling autonomous vehicles with embedded systems. In 2018 ACM/IEEE 9th International Conference on Cyber-Physical Systems 287–296 (IEEE, 2018).

  20. Cao, Z., Xu, S., Peng, H., Yang, D. & Zidek, R. Confidence-aware reinforcement learning for self-driving cars. IEEE Trans. Intell. Transp. Syst. 23, 7419–7430 (2022).

    Article  Google Scholar 

  21. Thomas, P. S. et al. Preventing undesirable behavior of intelligent machines. Science 366, 999–1004 (2019).

    Article  Google Scholar 

  22. Levine, S., Kumar, A., Tucker, G. & Fu, J. Offline reinforcement learning: tutorial, review, and perspectives on open problems. Preprint at arXiv (2020).

  23. Garcıa, J. & Fernández, F. A comprehensive survey on safe reinforcement learning. J. Mach. Learn. Res. 16, 1437–1480 (2015).

    MATH  MathSciNet  Google Scholar 

  24. Achiam, J., Held, D., Tamar, A. & Abbeel, P. Constrained policy optimization. In International Conference on Machine Learning 22–31 (JMLR, 2017).

  25. Berkenkamp, F., Turchetta, M., Schoellig, A. & Krause, A. Safe model-based reinforcement learning with stability guarantees. Adv. Neural Inf. Process. Syst. 30, 908-919 (2017).

  26. Ghadirzadeh, A., Maki, A., Kragic, D. & Björkman, M. Deep predictive policy training using reinforcement learning. In 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems 2351–2358 (IEEE, 2017).

  27. Abbeel, P. & Ng, A. Y. Apprenticeship learning via inverse reinforcement learning. In Proc. Twenty-first International Conference on Machine Learning, 1 (Association for Computing Machinery, 2004).

  28. Abbeel, P. & Ng, A. Y. Exploration and apprenticeship learning in reinforcement learning. In Proc. 22nd International Conference on Machine Learning 1–8 (Association for Computing Machinery, 2005).

  29. Ross, S., Gordon, G. & Bagnell, D. A reduction of imitation learning and structured prediction to no-regret online learning. In Gordon, G., Dunson, D. & Dudík, M. (eds) Proc. Fourteenth International Conference on Artificial Intelligence and Statistics, 627–635 (JMLR, 2011).

  30. Zhang, J. & Cho, K. Query-efficient imitation learning for end-to-end autonomous driving. In Thirty-First AAAI Conference on Artificial Intelligence (AAAI), 2891–2897 (AAAI Press, 2017).

  31. Bicer, Y., Alizadeh, A., Ure, N. K., Erdogan, A. & Kizilirmak, O. Sample efficient interactive end-to-end deep learning for self-driving cars with selective multi-class safe dataset aggregation. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems 2629–2634 (IEEE, 2019).

  32. Alshiekh, M. et al Safe reinforcement learning via shielding. In Proc. Thirty-Second AAAI Conference on Artificial Intelligence Vol. 32, 2669-2678 (AAAI Press, 2018).

  33. Brun, W., Keren, G., Kirkeboen, G. & Montgomery, H. Perspectives on Thinking, Judging, and Decision Making (Universitetsforlaget, 2011).

  34. Dabney, W. et al. A distributional code for value in dopamine-based reinforcement learning. Nature 577, 671–675 (2020).

    Article  Google Scholar 

  35. Cao, Z. et al. A geometry-driven car-following distance estimation algorithm robust to road slopes. Transp. Res. Part C 102, 274–288 (2019).

    Article  Google Scholar 

  36. Xu, S. et al. System and experiments of model-driven motion planning and control for autonomous vehicles. IEEE Trans. Syst. Man. Cybern. Syst. 52, 5975–5988 (2022).

    Article  Google Scholar 

  37. Cao, Z. Codes and data for dynamic confidence-aware reinforcement learning. DCARL. Zenodo (2022).

  38. Kochenderfer, M. J. Decision Making Under Uncertainty: Theory and Application (MIT Press, 2015).

  39. Ivanovic, B. et al. Heterogeneous-agent trajectory forecasting incorporating class uncertainty. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 12196–12203 (IEEE, 2022).

  40. Yang, Y., Zha, K., Chen, Y., Wang, H. & Katabi, D. Delving into deep imbalanced regression. In International Conference on Machine Learning 11842–11851 (PMLR, 2021).

  41. Efron, B. & Tibshirani, R. J. An Introduction to the Bootstrap (CRC Press, 1994).

  42. Dosovitskiy, A., Ros, G., Codevilla, F., Lopez, A. & Koltun, V. CARLA: An open urban driving simulator. In Proceedings of the 1st Annual Conference on Robot Learning, 1–16 (PMLR, 2017).

Download references


This work is supported by the National Natural Science Foundation of China (NSFC) (U1864203 (D.Y.), 52102460 (Z.C.), 61903220 (K.J.)), China Postdoctoral Science Foundation (2021M701883 (Z.C.)) and Beijing Municipal Science and Technology Commission (Z221100008122011 (D.Y.)). It is also funded by the Tsinghua University-Toyota Joint Center (D.Y.).

Author information

Authors and Affiliations



Z.C., S.X., D.Y. and H.P. developed the performance improvement technique, which can outperform the existing self-driving policy. Z.C. and W.Z. developed the continuous improvement technique using the worst confidence value. Z.C., S.X. and W.Z. designed the whole self-driving platform in the real world. Z.C., K.J. and D.Y. designed and conducted the experiments and collected the data.

Corresponding author

Correspondence to Diange Yang.

Ethics declarations

Competing interests

The authors declare no competing interests.

Peer review

Peer review information

Nature Machine Intelligence thanks Ali Alizadeh and the other, anonymous, reviewer(s) for their contribution to the peer review of this work.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary information

Supplementary Information

Essential supplementary description for the proposed technology, detailed setting and results for the experiments, descriptions of the data file and vehicles.

Supplementary Video 1

Evaluation results on running self-driving vehicle.

Supplementary Video 2

Continuous performance improvement using confidence value.

Supplementary Video 3

Comparing with classical value-based RL agent.

Supplementary Software

Software to run and render the results of experiments 1 to 3.

Rights and permissions

Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.

Reprints and Permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Cao, Z., Jiang, K., Zhou, W. et al. Continuous improvement of self-driving cars using dynamic confidence-aware reinforcement learning. Nat Mach Intell 5, 145–158 (2023).

Download citation

  • Received:

  • Accepted:

  • Published:

  • Issue Date:

  • DOI:


Quick links

Nature Briefing

Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

Get the most important science stories of the day, free in your inbox. Sign up for Nature Briefing