The Role of Reinforcement Learning in Autonomous Systems: Applications, Challenges, and Future Directions

1. INTRODUCTION

Autonomous systems are computational or physical systems that can perceive their surroundings, make decisions, and perform actions with limited continuous human intervention. Examples include mobile robots, autonomous vehicles, unmanned aerial vehicles, warehouse robots, and intelligent industrial systems. These systems operate through a sequence of decisions: a current observation influences an action, the action changes the state of the environment, and the new state affects later decisions. Consequently, autonomy is not simply a problem of predicting an output; it is also a problem of selecting actions whose consequences may extend over time.

Reinforcement learning addresses this type of sequential decision-making by allowing an agent to learn from interaction with an environment. Instead of receiving a correct labelled answer for every state, the agent receives feedback in the form of rewards and learns a policy that seeks to maximize cumulative return. Sutton and Barto describe RL as a framework for learning from interaction, where the central problem is the selection of actions that improve long-term reward¹.

The connection between RL and autonomous systems is therefore natural. A robot navigating a building, a vehicle responding to traffic situations, or a drone exploring an unfamiliar area must repeatedly choose actions under changing conditions. RL provides methods for learning such decision policies, while related fields such as perception, planning, control, and safety engineering provide additional components required for a complete autonomous system. Research in robotics, deep RL, autonomous driving, and mobile-robot exploration has consequently examined RL as a tool for learning control and decision-making policies².

This paper is a literature-focused research study. It does not present the student’s separate coding project as an experimental result. Instead, it synthesizes established research to explain where RL is being applied, what methods are used, what limitations repeatedly appear in the literature, and where further research is needed.

2. PURPOSE OF THE STUDY

The purpose of this study is to examine the role of reinforcement learning in autonomous systems and to develop a clear understanding of its applications, methods, limitations, and future directions. The study is designed to connect the mathematical foundations of RL with real autonomous-system problems while maintaining a distinction between established research findings and the student’s separate practical implementation.

3. OBJECTIVES

The objectives of the study are:

1. to explain the fundamental concepts of reinforcement learning and their relationship with autonomous decision-making;

2. to review important applications of RL in robotics, autonomous navigation, vehicles, drones, and intelligent control;

3. to compare major RL approaches that are relevant to autonomous systems;

4. to identify recurring technical and deployment challenges reported in the literature;

5. to identify research gaps that remain important for reliable autonomous decision-making; and

6. to discuss future directions for safe, efficient, robust, and explainable RL-based autonomy.

4. RESEARCH QUESTIONS

The review is guided by the following research questions:

RQ1: How does reinforcement learning support sequential decision-making in autonomous systems?

RQ2: Which RL methods are commonly relevant to autonomous robotics, navigation, vehicles, and drones?

RQ3: What challenges limit the reliable deployment of RL in physical autonomous systems?

RQ4: What research directions could improve the safety, efficiency, robustness, and interpretability of RL-based autonomy?

5. METHODOLOGY

This study follows a focused literature-review methodology. The objective is to synthesize foundational RL knowledge and research that specifically connects RL with autonomous systems. The review is organized around four stages: defining the research questions, selecting relevant academic and institutional sources, classifying the literature by application and method, and synthesizing recurring findings, challenges, and research gaps.

5.1 Literature Selection Approach

The sources considered for this paper include foundational RL literature, peer-reviewed surveys and research papers on robotics and autonomous driving, work on mobile-robot exploration, research on sim-to-real transfer, literature on safe RL, and an established AI risk-management framework. Priority is given to sources that directly explain RL methods or analyze their use in autonomous systems. The paper uses the selected sources to identify recurring themes rather than claiming to be a systematic meta-analysis of every publication in the field.

5.2 Analysis Framework

The reviewed literature is organized into the following themes:

• RL fundamentals and decision-making;
• robotics and autonomous navigation;
• autonomous driving and vehicle control;
• drones and mobile-robot exploration;
• deep and value-based RL methods;
• safety, robustness, and trustworthy deployment;
• simulation and sim-to-real transfer; and
• research gaps and future directions.

Figure 1: Literature-review workflow used to organize the study.

6. BACKGROUND AND CONTEXT

6.1 Reinforcement Learning Fundamentals

In reinforcement learning, an agent interacts with an environment over a sequence of time steps. At time t, the agent observes a state s_t, selects an action a_t, receives a reward r_t+1, and transitions to a new state s_t+1. The objective is to learn a policy that maximizes expected cumulative reward.

A common mathematical representation is the Markov Decision Process (MDP), represented by (S, A, P, R, γ), where S denotes the state space, A the action space, P the transition model, R the reward function, and γ the discount factor. The discounted return is

(1)

The discount factor determines how strongly future rewards contribute to the value of a current decision³.

Figure 2: Basic reinforcement-learning interaction loop used to explain autonomous decision-making.

6.2 Exploration and Exploitation

An autonomous agent must balance exploration and exploitation. Exploration means trying actions that may reveal a better strategy, while exploitation means choosing an action that currently appears useful. An ϵ-greedy strategy is a common approach: with probability ϵ the agent explores, and otherwise it selects an action with the highest estimated value. This balance is particularly important in autonomous navigation because learning requires discovering alternatives, while deployment requires predictable behaviour.

6.3 Q-Learning

Q-learning is a model-free, value-based RL algorithm that estimates the long-term usefulness of taking action a in state s. Its update rule is

(2)

where α is the learning rate and γ is the discount factor. The update combines the current estimate with immediate reward and the estimated future value of the next state. Q-learning is especially transparent for small discrete environments because the learned values can be inspected directly.

6.4 Deep Reinforcement Learning

When the state space is too large for a table, neural networks can be used to approximate value functions or policies. Mnih et al. demonstrated a Deep Q-Network (DQN) that learned from high-dimensional sensory inputs in a reinforcement-learning setting⁴. Deep RL has subsequently become important in research involving complex perception and decision-making, including autonomous driving and robotics⁵.

Figure 3: Simplified taxonomy of major RL method families discussed in the literature.

7. RELATED WORK AND LITERATURE REVIEW

Sutton and Barto provide the foundational framework for understanding RL as learning through interaction and cumulative reward⁶. Their formulation explains why RL differs from conventional supervised learning: the agent is not simply learning from a fixed collection of labelled input-output pairs, but is learning how current actions influence future states and rewards.

Kober, Bagnell, and Peters surveyed reinforcement learning in robotics and described its relevance to learning control policies and motor skills. Their work also emphasized practical difficulties such as exploration, high-dimensional state spaces, and the cost of experimentation on physical robots⁷. Kober and Peters further discussed the transition from algorithms to robot experiments, highlighting the relationship between learning methods and physical control⁸.

Mnih et al. introduced the DQN approach and demonstrated learning from high-dimensional sensory observations in a reinforcement-learning setting⁹. Although the original work used game environments rather than autonomous vehicles, its significance for autonomy is methodological: neural networks can be used to approximate action values when direct tabular representation is not practical.

Pan et al. reviewed reinforcement-learning-based robotics research and identified applications across navigation, manipulation, control, and related robotic tasks¹⁰. Kiran et al. reviewed deep RL for autonomous driving and discussed its use in decision-making and control while identifying issues involving validation, simulation, data requirements, and safety¹¹. Martin et al. reviewed RL for mobile-robot exploration and highlighted recurring issues involving exploration–exploitation, reward shaping, state-space complexity, and convergence¹².

The transfer from simulation to physical systems is another major theme. Zhao, Queralta, and Westerlund reviewed sim-to-real transfer for deep RL in robotics and explained why simulation is attractive for scalable training while differences between simulation and reality can reduce the effectiveness of a learned policy after transfer¹³. More recent safe-RL research treats safety as a distinct methodological concern, particularly where learning actions can have physical consequences¹⁴.

8. APPLICATIONS OF REINFORCEMENT LEARNING IN AUTONOMOUS SYSTEMS

8.1 Robotics

Robotics is one of the major application areas for RL. A robot can learn a policy for movement, manipulation, navigation, or control by repeatedly interacting with its environment. The literature describes RL as a useful approach when the desired behaviour is difficult to specify entirely through hand-written rules¹⁵.

Table 1: Selected Literature Themes Relevant to RL-Based Autonomy

Table 2: RL Application Landscape in Autonomous Systems

8.2 Autonomous Navigation

Navigation is a direct example of sequential decision-making. An autonomous agent must select a sequence of actions that moves it toward a destination while responding to obstacles and changes in the environment. RL can represent this problem through states, actions, rewards, and a learned policy. Research on mobile robots has investigated RL for exploration and navigation, including the balance between discovering new information and exploiting known routes¹⁶.

8.3 Autonomous Vehicles

Autonomous driving involves a collection of decisions, including behaviour selection, motion planning, and control. Survey literature has examined RL and deep RL for these tasks while also emphasizing the difficulty of validating learned behaviour in safety-critical conditions¹⁷. The role of RL is therefore not limited to one component of a vehicle; it can be studied at different levels of the decision and control pipeline.

8.4 Drones and Unmanned Aerial Vehicles

Drones operate in three-dimensional environments and may need to navigate, explore, avoid obstacles, and respond to changing conditions. RL is relevant because the agent can learn a policy for selecting actions based on the observed state. The same fundamental trade-offs found in other autonomous systems—exploration, reward design, computational requirements, and safety—also apply to aerial systems.

8.5 Industrial and Intelligent Control

Industrial systems often require repeated control decisions under operational constraints. RL can be investigated for optimization and adaptive control where the system can be represented as a sequential decision problem. In such settings, however, safety and operational constraints are essential because unrestricted trial-and-error behaviour may not be acceptable in a production environment.

9. RL METHODS RELEVANT TO AUTONOMOUS SYSTEMS

Different RL methods are appropriate for different state and action spaces. Tabular methods are useful for small discrete problems, while function approximation and policy-optimization methods become more relevant as the problem becomes larger or more continuous.

Table 3: Common RL Approaches and Their Characteristics

9.1 Q-Learning as a Baseline

Q-learning is attractive as a baseline because its update mechanism is mathematically simple and its learned values can be inspected. For a small discrete navigation environment, a Q-table can show which actions the agent has learned to prefer in each state. This transparency makes Q-learning particularly useful for educational demonstrations and controlled experiments, even though larger autonomous systems may require more advanced methods.

9.2 Deep and Policy-Based Methods

DQN extends value-based learning through neural-network approximation. Policy-gradient and actor–critic methods provide alternatives when the policy itself or continuous control actions are central to the problem. The choice of algorithm depends on the state representation, action space, safety requirements, computational resources, and evaluation conditions rather than on one universally applicable method.

10. KEY FINDINGS FROM THE LITERATURE

10.1 RL Fits Sequential Decision-Making

A recurring finding is that RL is well aligned with problems in which actions influence future states and rewards. This is a fundamental property of autonomous behaviour because a decision made at one moment can affect the options available later¹⁸.

10.2 RL Supports Adaptive Behaviour

Unlike a fixed rule set, an RL policy can be learned from interaction. This creates the possibility of adapting behaviour to a task environment, although the quality of the adaptation depends on the reward function, training experience, and environment representation¹⁹.

10.3 Deep RL Expands the Range of Problems

Deep RL extends RL to larger state representations through neural networks. This is particularly relevant when an autonomous system receives rich sensory information rather than a small set of discrete variables²⁰.

10.4 Safety Is a Core Deployment Issue

The literature indicates that high performance in a simulated environment is not sufficient evidence for safe physical deployment. Safe-RL research explicitly addresses the challenge of learning while respecting constraints, while sim-to-real research highlights differences between simulated and real environments²¹.

10.5 Evaluation Must Be Multi-Dimensional

Autonomous behaviour should not be represented by a single performance number. Depending on the application, evaluation can include task success, cumulative reward, path efficiency, collisions, robustness, computational cost, and behaviour under changed environmental conditions. A multi-dimensional evaluation provides a more informative description of an autonomous policy.

Figure 4: Key challenge dimensions that influence reliable deployment of RL-based autonomous systems.

11. CHALLENGES

11.1 Reward Design

The reward function defines what behaviour the agent is encouraged to learn. A poorly designed reward may produce behaviour that improves the numerical objective without satisfying the intended task. In autonomous systems, this can become especially important when a shortcut in the reward function conflicts with safety or operational requirements.

11.2 Sample Efficiency and Training Cost

RL agents may require many interactions before learning useful behaviour. Real-world interaction can be expensive, time-consuming, or unsafe, which is one reason simulation is widely used for development. However, the use of simulation introduces modelling assumptions and can create a transfer problem when the learned policy is moved to a physical system²².

11.3 Exploration and Safety

Exploration is necessary for learning, but unrestricted exploration is undesirable when incorrect actions can damage equipment or endanger people. Safe RL therefore studies ways to incorporate constraints, risk measures, or safety mechanisms into learning and deployment²³.

11.4 Generalization

A policy trained in one environment may not automatically perform well when the environment changes. Changes in obstacle positions, traffic patterns, sensor conditions, or physical dynamics can expose weaknesses in a learned policy. Generalization is therefore important for systems expected to operate beyond the exact situations encountered during training.

11.5 Sim-to-Real Transfer

Simulation can provide large amounts of experience without exposing physical hardware to repeated failures. Nevertheless, simulation cannot perfectly reproduce sensor noise, friction, latency, dynamics, lighting, and other real-world effects. Sim-to-real research therefore investigates transfer strategies that reduce performance loss when a policy moves from simulation to physical systems²⁴.

11.6 Interpretability and Trust

Complex autonomous systems may need to explain or justify decisions to engineers, operators, regulators, or users. NIST’s AI Risk Management Framework emphasizes systematic consideration of risks and trustworthy-AI characteristics throughout the AI lifecycle²⁵. For RL-based autonomy, this creates a need for evaluation methods that go beyond task performance and also consider transparency, reliability, and risk.

12. RESEARCH GAPS

The reviewed literature points to several areas where further work remains important.

1. Reliable safety mechanisms: autonomous RL systems need methods that can limit unsafe exploration and provide evidence of safe behaviour before deployment.

2. Efficient learning: reducing the number of interactions needed to learn useful policies remains important, especially when real-world data collection is expensive.

3. Robust generalization: policies should remain useful when environments, sensor conditions, or task parameters change.

4. Improved sim-to-real transfer: simulation should better support the transition to physical systems without large performance losses.

5. Explainable decision-making: users and engineers may need clearer information about why a learned policy selected a particular action.

6. Multi-agent coordination: future autonomous environments may contain many interacting agents, creating additional coordination and communication challenges.

These gaps are not independent. For example, improving sample efficiency may reduce the need for extensive real-world exploration, while better sim-to-real transfer can make simulated training more useful. Similarly, explainability and safety can reinforce each other by making policy behaviour easier to inspect and validate.

13. DISCUSSION

The literature reviewed in this paper supports the view that RL is particularly relevant when an autonomous system must repeatedly choose actions and learn how those actions influence future states. Its central strength is the ability to optimize long-term behaviour through feedback rather than relying only on manually specified responses.

At the same time, RL should not be interpreted as a complete replacement for sensing, perception, planning, control, or safety mechanisms. In practical autonomous systems, an RL policy can form one component of a larger architecture. The appropriate design depends on the application, available data, safety requirements, computational resources, and level of autonomy required.

A useful way to summarize the relationship is shown below:

Figure 5: Conceptual position of an RL policy within a larger autonomous-system architecture.

Table 4: Comparison of Common Decision-Making Approaches

14. FUTURE RESEARCH DIRECTIONS

Future research can develop RL-based autonomy along several complementary directions.

14.1 Safe Reinforcement Learning

Future systems can incorporate safety constraints directly into the learning process rather than treating safety as an external check. This is particularly important for physical autonomous systems where unsafe exploration may not be acceptable²⁶.

14.2 Sim-to-Real Learning

More robust transfer techniques can make simulation-based training more representative of physical deployment. Domain randomization, adaptation, system identification, and related approaches can be studied to reduce the gap between simulated and real environments²⁷.

14.3 Multi-Agent Reinforcement Learning

Many future autonomous environments will contain multiple interacting agents, such as vehicle traffic, robot fleets, or groups of drones. Multi-agent RL can study coordination, communication, competition, and cooperation among such agents.

14.4 Explainable and Human-Aware RL

Future systems can provide clearer explanations of policy decisions and incorporate human preferences or constraints into the learning process. This direction is relevant to trust, oversight, and responsible deployment.

14.5 Hybrid Learning and Control

Combining RL with established planning and control methods may provide a practical balance between adaptive learning and explicit constraints. Such hybrid architectures can be investigated for applications where reliability is as important as adaptability.

15. LIMITATIONS OF THE STUDY

This paper has several limitations. First, it is a focused literature review rather than a systematic meta-analysis of every publication in the field. Second, the selected literature covers several autonomous-system domains, so the discussion is intentionally broad rather than deeply specialized in one platform. Third, the paper does not claim that a small student implementation can represent the complexity of a real autonomous vehicle, robot, or drone. The separate practical project is therefore treated as an educational implementation and is not presented as empirical evidence for the literature findings.

16. CONCLUSION

Reinforcement learning provides a useful framework for autonomous systems because it addresses sequential decision-making through interaction, feedback, and policy improvement. The literature shows applications across robotics, autonomous navigation, autonomous vehicles, drones, exploration, and intelligent control. At the same time, real-world deployment requires more than successful learning in a controlled environment. Reward design, sample efficiency, safety, generalization, computational requirements, interpretability, and sim-to-real transfer remain important considerations.

The central conclusion of this review is that the role of RL in autonomy is best understood as part of a broader decision-making and control architecture. RL can provide adaptive policies, but reliable autonomous behaviour also depends on perception, planning, control, validation, safety mechanisms, and appropriate human oversight. Future research should therefore focus not only on improving learning algorithms but also on making learned policies more efficient, robust, safe, explainable, and transferable to real environments.

ACKNOWLEDGEMENT

The author expresses sincere gratitude to Assistant Professor Smita Mangesh Junnarkar, Department of Data Science, SIES College of Arts, Science and Commerce (Empowered Autonomous), Sion (W), Mumbai, for her guidance and academic support during the development of this research work.

Endnotes

1. R. S. Sutton and A. G. Barto, Reinforcement Learning: An Introduction, 2nd ed. Cambridge, MA, USA: MIT Press, 2018.

2. J. Kober, J. A. Bagnell, and J. Peters, “Reinforcement learning in robotics: A survey,” International Journal of Robotics Research, vol. 32, no. 11, pp. 1238–1274, 2013, https://doi.org/10.1177/0278364913495721.; V. Mnih et al., “Human-level control through deep reinforcement learning,” Nature, vol. 518, pp. 529–533, 2015, https://doi.org/10.1038/nature14236.; Z. Pan et al., “A systematic review on reinforcement learning-based robotics within the last decade,” IEEE Access, vol. 8, pp. 176598–176623, 2020, https://doi.org/10.1109/ACCESS.2020.3027152.; B. R. Kiran et al., “Deep reinforcement learning for autonomous driving: A survey,” IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 6, pp. 4909–4926, 2022, https://doi.org/10.1109/TITS.2021.3054625.; J. Martin et al., “Reinforcement learning for mobile robotics exploration: A survey,” IEEE Transactions on Neural Networks and Learning Systems, vol. 34, no. 8, pp. 3796–3810, 2023, https://doi.org/10.1109/TNNLS.2021.3124466.

3. R. S. Sutton and A. G. Barto, Reinforcement Learning: An Introduction, 2nd ed. Cambridge, MA, USA: MIT Press, 2018.

4. V. Mnih et al., “Human-level control through deep reinforcement learning,” Nature, vol. 518, pp. 529–533, 2015, https://doi.org/10.1038/nature14236.

5. B. R. Kiran et al., “Deep reinforcement learning for autonomous driving: A survey,” IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 6, pp. 4909–4926, 2022, https://doi.org/10.1109/TITS.2021.3054625.; H. Li et al., “A survey on recent advancements in autonomous driving using deep reinforcement learning: Applications, challenges, and solutions,” IEEE Transactions on Intelligent Transportation Systems, vol. 25, no. 12, pp. 19365–19398, 2024, https://doi.org/10.1109/TITS.2024.3452480.

6. R. S. Sutton and A. G. Barto, Reinforcement Learning: An Introduction, 2nd ed. Cambridge, MA, USA: MIT Press, 2018.

7. J. Kober, J. A. Bagnell, and J. Peters, “Reinforcement learning in robotics: A survey,” International Journal of Robotics Research, vol. 32, no. 11, pp. 1238–1274, 2013, https://doi.org/10.1177/0278364913495721.

8. J. Kober and J. Peters, Learning Motor Skills: From Algorithms to Robot Experiments. Cham, Switzerland: Springer, 2014, https://doi.org/10.1007/978-3-319-03194-1.

9. V. Mnih et al., “Human-level control through deep reinforcement learning,” Nature, vol. 518, pp. 529–533, 2015, https://doi.org/10.1038/nature14236.

10. Z. Pan et al., “A systematic review on reinforcement learning-based robotics within the last decade,” IEEE Access, vol. 8, pp. 176598–176623, 2020, https://doi.org/10.1109/ACCESS.2020.3027152.

11. B. R. Kiran et al., “Deep reinforcement learning for autonomous driving: A survey,” IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 6, pp. 4909–4926, 2022, https://doi.org/10.1109/TITS.2021.3054625.

12. J. Martin et al., “Reinforcement learning for mobile robotics exploration: A survey,” IEEE Transactions on Neural Networks and Learning Systems, vol. 34, no. 8, pp. 3796–3810, 2023, https://doi.org/10.1109/TNNLS.2021.3124466.

13. W. Zhao, J. P. Queralta, and T. Westerlund, “Sim-to-real transfer in deep reinforcement learning for robotics: A survey,” arXiv:2009.13303, 2020.

14. Z. Zhou et al., “A review of safe reinforcement learning: Methods, theories, and applications,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 12, pp. 11216–11235, 2024, https://doi.org/10.1109/TPAMI.2024.3457538.

15. J. Kober, J. A. Bagnell, and J. Peters, “Reinforcement learning in robotics: A survey,” International Journal of Robotics Research, vol. 32, no. 11, pp. 1238–1274, 2013, https://doi.org/10.1177/0278364913495721.; Z. Pan et al., “A systematic review on reinforcement learning-based robotics within the last decade,” IEEE Access, vol. 8, pp. 176598–176623, 2020, https://doi.org/10.1109/ACCESS.2020.3027152.

16. J. Martin et al., “Reinforcement learning for mobile robotics exploration: A survey,” IEEE Transactions on Neural Networks and Learning Systems, vol. 34, no. 8, pp. 3796–3810, 2023, https://doi.org/10.1109/TNNLS.2021.3124466.

17. B. R. Kiran et al., “Deep reinforcement learning for autonomous driving: A survey,” IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 6, pp. 4909–4926, 2022, https://doi.org/10.1109/TITS.2021.3054625.; H. Li et al., “A survey on recent advancements in autonomous driving using deep reinforcement learning: Applications, challenges, and solutions,” IEEE Transactions on Intelligent Transportation Systems, vol. 25, no. 12, pp. 19365–19398, 2024, https://doi.org/10.1109/TITS.2024.3452480.

18. R. S. Sutton and A. G. Barto, Reinforcement Learning: An Introduction, 2nd ed. Cambridge, MA, USA: MIT Press, 2018.

19. R. S. Sutton and A. G. Barto, Reinforcement Learning: An Introduction, 2nd ed. Cambridge, MA, USA: MIT Press, 2018.; V. Mnih et al., “Human-level control through deep reinforcement learning,” Nature, vol. 518, pp. 529–533, 2015, https://doi.org/10.1038/nature14236.

20. V. Mnih et al., “Human-level control through deep reinforcement learning,” Nature, vol. 518, pp. 529–533, 2015, https://doi.org/10.1038/nature14236.; B. R. Kiran et al., “Deep reinforcement learning for autonomous driving: A survey,” IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 6, pp. 4909–4926, 2022, https://doi.org/10.1109/TITS.2021.3054625.

21. Z. Zhou et al., “A review of safe reinforcement learning: Methods, theories, and applications,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 12, pp. 11216–11235, 2024, https://doi.org/10.1109/TPAMI.2024.3457538.; W. Zhao, J. P. Queralta, and T. Westerlund, “Sim-to-real transfer in deep reinforcement learning for robotics: A survey,” arXiv:2009.13303, 2020.

22. W. Zhao, J. P. Queralta, and T. Westerlund, “Sim-to-real transfer in deep reinforcement learning for robotics: A survey,” arXiv:2009.13303, 2020.

23. Z. Zhou et al., “A review of safe reinforcement learning: Methods, theories, and applications,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 12, pp. 11216–11235, 2024, https://doi.org/10.1109/TPAMI.2024.3457538.

24. W. Zhao, J. P. Queralta, and T. Westerlund, “Sim-to-real transfer in deep reinforcement learning for robotics: A survey,” arXiv:2009.13303, 2020.

25. E. Tabassi, “Artificial Intelligence Risk Management Framework (AI RMF 1.0),” National Institute of Standards and Technology, NIST AI 100-1, 2023, https://doi.org/10.6028/NIST.AI.100-1.

26. Z. Zhou et al., “A review of safe reinforcement learning: Methods, theories, and applications,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 12, pp. 11216–11235, 2024, https://doi.org/10.1109/TPAMI.2024.3457538.

27. W. Zhao, J. P. Queralta, and T. Westerlund, “Sim-to-real transfer in deep reinforcement learning for robotics: A survey,” arXiv:2009.13303, 2020.