The State of Reinforcement Learning in 2025

Introduction

Machine Learning broadly encompasses three categories: Supervised, Unsupervised, and Reinforcement Learning. Reinforcement Learning (RL) is a type of ML in which an agent learns how to make decisions by interacting with an environment to achieve a specific goal. It replicates the trial-and-error learning process humans use to accomplish their goals.

The agent's objective is to maximize a cumulative reward over time. Unlike supervised learning, RL does not rely on a training dataset but learns from feedback that evaluates performance without predefined behavioral targets. This dynamic learning process has enabled RL to excel in areas requiring sequential decision-making, from robotics to financial modeling.

In essence, RL enables the creation of intelligent agents, which are computer programs capable of making decisions. Reinforcement learning, similar to how humans learn, is especially effective in uncertain and complex environments.

Types and Applications of Reinforcement Learning by Industry

RL is broadly categorized into three main types: value-based, policy-based, and model-based methods.

• Value-based approaches, such as Q-learning, focus on estimating the value of actions to determine the best policy indirectly.
• Policy-based methods, like Policy Gradient, directly optimize the policy itself, making them suitable for high-dimensional action spaces.
• Model-based RL incorporates an internal model of the environment, enabling agents to plan and simulate outcomes before acting.

These diverse methodologies equip RL with the flexibility to tackle a wide range of real-world problems. For instance, deep reinforcement learning (DRL) has advanced the field by integrating deep learning techniques with RL algorithms. Leveraging the representational power of neural networks, DRL allows agents to process high-dimensional input data, such as images and complex sensor readings. Techniques like Deep Q-Networks (DQN) and Proximal Policy Optimization (PPO) have demonstrated exceptional performance in environments as varied as video games and autonomous driving.

Prominent companies are already leveraging RL to power products and services. To give you some examples of RL applications already available today:

• Tesla employs RL in its autopilot system to enable real-time decision-making for autonomous vehicles.
• SpaceX has used RL to improve the precision of rocket landings, a critical component of its reusable rocket technology. A SpaceX Super Heavy booster rocket successfully returned to Earth and was caught by giant robotic arms, marking the first attempt to recover the 232-foot booster at the launch tower after supporting the reusable Starship spacecraft's launch.
• In gaming, companies like DeepMind have utilized RL to create AI agents that outperform human players in complex games like StarCraft II.
• ChatGPT uses Reinforcement Learning from Human Feedback (RLHF) to align its responses with user preferences. After pretraining on vast datasets, RLHF refines the model by leveraging human feedback to train a reward model and optimize response quality through reinforcement learning techniques like Proximal Policy Optimization (PPO).
• Your smart robot vacuum avoids obstacles by receiving feedback from its environment and adapting its behavior to prevent collisions with walls, furniture, or stairs.

These examples underscore RL's impact, positioning it at the heart of future technological progress.

Rising stars that use RL to power their products

Based on public information gathered from Crunchbase, Pitchbook, and other open sources, we have compiled a list of venture-backed startups that leverage RL to transform their respective industries with ultimately innovative products. Check out the innovative companies that constitute today's Innovation Landscape in AI in Reinforcement Learning in early 2025.

Let's look at each category one by one.

Advanced Robotics

In Robotics, RL enables machines to adapt and optimize their performance in dynamic environments. Research and practical applications span various robotic domains, such as quadruped locomotion, drone navigation, wheeled robotics, and object manipulation. For instance, companies like Swiss-Mile focus on quadruped locomotion, while Shearwater AI and ANDRO Innovation Lab contribute to drone navigation with advanced RL-powered software and recognition by platforms like Tradewinds Solutions Marketplace.

In wheeled robotics, Unbox Robotics applies RL to automate and optimize movement, reminiscent of consumer robots like Roombas. Similarly, companies like Covariant and Osaro lead in object manipulation through sophisticated RL-driven software solutions.

Physical Intelligence has further advanced RL in robotics by developing the world's first generalist policy for robotic hands, capable of executing tasks based on textual or voice instructions. Unlike competitors, their models are trained to perform a wide array of tasks, showcasing versatility and adaptability. While Skild.ai emerges as a direct competitor, it currently lacks a generalist policy. Open-source initiatives like OmniDrones provide RL training platforms based on NVIDIA's Isaac Sim, fostering innovation in drone control.

Autonomous Driving

RL is driving innovation in Autonomous Driving enabling systems to optimize decision-making through trial and error in dynamic, real-world environments.

Companies such as Wayve, which recently launched a test campaign in California in collaboration with Uber, and Waymo, Alphabet's subsidiary, are at the forefront of this field, attracting substantial investments. Established automotive giants like Tesla, Audi, BMW, and Ford also leverage RL to enhance their autonomous driving capabilities, often outperforming traditional rule-based approaches.

Despite these advancements, the inability to explain the decisions of RL models — a challenge tied to the broader issue of explainable AI — has raised concerns about trust and accountability in these systems.

In the broader context of smart cities, RL is instrumental in optimizing traffic flow, reducing congestion, and improving urban mobility. Autonomous vehicles are central to this vision, requiring RL methods that ensure safety and reliability in complex environments.

AI Research & Development

In AI R&D, RL is extensively used to enhance software functionality across fields.

Axiomatic AI has developed the Automated Interpretable Reasoning (AIR) model, which integrates reinforcement learning, large language models (LLMs), and world models to automate and enhance prototype development, particularly in semiconductor hardware. AIR enables AI systems to learn optimal decision-making strategies through trial and error, improving efficiency and innovation in engineering and scientific research.

Another advanced use case involves RL-trained generative models tailored to write code resembling a company's specific style, enhancing software development workflows. Poolside is developing advanced AI models tailored for software engineering, utilizing a novel approach called Reinforcement Learning from Code Execution Feedback (RLCEF) to enhance code generation and reasoning capabilities. Their flagship model, Malibu, is trained specifically to address complex software engineering challenges, aiming to assist developers in building software more efficiently and effectively.

Defense

RL is increasingly applied in Defense to automate critical and high-risk tasks, reducing reliance on personnel whose loss is costly. Shield AI, for instance, aims to develop a "Hivemind" system that enables aerial vehicles to operate autonomously without GPS, communication, or pilots. While the company remains vague about its specific use of RL, its mission to enable autonomous mission execution aligns with RL applications. Shield AI's valuation has approached $1B, bolstered by a recent $200M investment round.

Similarly, Anduril Industries focuses on automating tasks like drone mission management, though their explicit use of RL remains unconfirmed. Both companies highlight RL's potential to revolutionize operational efficiency and autonomy in defense applications.

Other notable use cases include Cohere Technology Group's anomaly detection system, which employs RL agents to identify network threats missed by traditional cybersecurity measures and suggest counteractions. In military strategy, RL is explored for war games, particularly through Hierarchical Reinforcement Learning (HRL), which structures decision-making processes to mirror modern organizational hierarchies. A recent dissertation highlights HRL's potential to enhance precision in war game simulations. However, current results remain modest, attributed to limited attempts at practical implementation.

Logistics

In Logistics RL is enhancing warehouse operations, planning, and process optimization. Companies like AICA, Covariant, and Osaro focus on deploying RL-powered robots for automated tasks in warehouses, improving efficiency and reducing manual labor. RL also aids in optimizing logistical planning. For example, DeepVu applies RL for general KPI optimization, Minds.ai leverages it in the semiconductor industry, and NNAISENSE uses RL with digital twins to simulate and refine logistics workflows. Another innovative concept involves RL-based adaptation of product packaging, such as Autoboxing, although this area remains in early development with limited investment.

Energy Management

RL is being applied in Energy Management to address inefficiencies in modern energy systems, which are often large and cumbersome. A notable example is EnliteAI, which has developed an RL-based energy management system capable of optimizing electricity distribution. The system not only improves energy allocation but also suggests solutions during emergency situations. Additionally, it facilitates network maintenance planning, enabling more efficient scheduling of repairs. A demo of their system is available online, showcasing its practical applications in managing energy grids.

However, further advancements in RL for energy management face limitations due to a lack of high-quality simulations that accurately reflect real-world scenarios constrains the ability to test and refine RL models comprehensively.

Agriculture

In Agriculture, RL is optimizing resource use, improving crop yields, and automating tasks. Research suggests RL can help manage water and fertilizer use more efficiently, detect pests and diseases using drones, and optimize planting patterns to enhance productivity. For example, RL-powered drones have been trained to adjust their velocity and height to apply precise amounts of pesticides. There is also exploration of RL-driven robots capable of harvesting crops like apples or pruning grapevines, although these technologies remain largely experimental with limited commercial use.

A typical RL workflow in Agriculture involves collecting data, creating a digital twin, training RL models in simulated environments, and deploying them in the real world. However, the absence of mature digital twin systems specific to agriculture is a major challenge. Current efforts are focused on building these simulations to enable RL applications, such as greenhouse energy management. Despite the potential, decision-making systems for plant care are still in the research phase due to limited data and the high cost of errors. While RL shows significant promise in advancing agricultural efficiency and sustainability, practical implementation remains constrained by technological and data limitations.

Greeneye Technology, for example, utilizes AI and deep learning to revolutionize agricultural pest control by enabling precise, selective spraying of herbicides. Their proprietary Selective Spraying (SSP) system integrates seamlessly with existing agricultural sprayers, allowing real-time identification and targeted application to weeds, which can reduce herbicide usage by up to 90%. This approach enhances crop management efficiency, reduces environmental impact, and increases profitability for farmers.

Taranis employs advanced AI and machine learning techniques to provide high-resolution, leaf-level imagery and actionable insights for crop management. Their platform captures detailed images of crops, enabling the detection of issues such as diseases, pests, and nutrient deficiencies. While Taranis utilizes sophisticated AI and machine learning algorithms, there is no explicit information indicating the use of reinforcement learning in their agricultural solutions.

Manufacturing

Additionally, RL is making strides in Manufacturing, i.e. in automated design and experimentation. Quilter employs RL for automating printed circuit board (PCB) designs, streamlining production in electronics manufacturing. In quantum physics, companies like Qruise integrate RL to create digital twins that assist researchers in experiments, helping uncover the parameters of quantum devices. These advancements highlight RL's versatility, from logistics to complex scientific research, underscoring its transformative potential across industries.

Biotech

In Biotech, RL has the potential to address critical challenges, such as workforce shortages in healthcare and the unpredictability of biotech research. Two primary use cases have emerged: dynamic treatment regimes and drug discovery assistance. While the first application, dynamic treatment regimes, is still poorly defined and underexplored among startups, existing research suggests RL could help optimize personalized treatment plans over time (e.g., through mathematical methods as discussed in referenced papers). However, the complexity of medical decision-making and limited understanding of the field poses challenges to broader adoption.

The second use case, drug discovery assistance, has gained more traction, with companies like Biomonadic, Ordaos, and Isomorphic Labs leading the way. These firms focus on software that accelerates and improves drug research processes. For example, Biomonadic and Ordaos specialize in optimizing research related to mini-proteins and plasmids, enhancing both speed and efficiency. Isomorphic Labs, supported by Alphabet and employing a sizable team of 164 professionals, is less explicit about its specific contributions but shows promise in pushing innovation forward. RL in biotech is helping to enhance research reliability, streamline processes, and potentially revolutionize drug discovery.

Fintech

In Fintech, RL is primarily applied to stock market prediction and high-frequency trading (HFT), although its efficacy and limitations spark skepticism. Most current RL-based solutions focus on short-term trading optimizations rather than long-term value investing. Notable examples include AI Capital Management, a hedge fund leveraging RL for HFT, and Equilibre Technologies, which combines RL with game theory for algorithmic trading. Equilibre's team, consisting of former Google DeepMind professionals, recently raised $7 million in funding, signaling credibility and potential for future innovation despite limited public details about their work.

Another interesting application of RL in fintech involves optimizing venture capital investments. Recent research introduced an RL agent designed to predict investment amounts in startups based on specific factors. While the model outperforms existing alternatives, the dataset's quality and evaluation metrics remain limitations. Researchers acknowledge that financial performance data of portfolio companies is absent, hindering predictive accuracy. Plans for future work include integrating richer datasets and refining evaluation metrics to better assess fund allocation. This underscores RL's potential in fintech and also highlights the need for robust data and methodological improvements for effective deployment.

Game Development

RL's role in Gaming is as a training ground for models and as a tool for game development. RL has powered breakthroughs in complex games, such as AlphaGo, OpenAI Five (DOTA2), and AlphaStar (StarCraft II), showcasing its ability to handle intricate strategies and dynamics. In game development, RL is used to create adaptive non-player characters (NPCs) and dynamic environments, with companies like rct AI leading in this space. Additionally, RL supports innovations in animation generation, such as Latent Technology's pre-seed efforts, and emerging areas like text-to-3D model creation, as seen with Irreverent Labs.

Education

RL is being explored in Education for applications like personalizing curricula, providing adaptive hints and quizzes, and optimizing A/B testing in educational platforms. RL is also used to model human students and generate educational content. However, these efforts face significant challenges, including unreliable and limited datasets, as well as uncertainties about the appropriate level of personalization for each student. There are also concerns about bias in the design of RL agents and the potential unintended consequences of their recommendations. While the technology remains in its infancy with no major startups yet, it holds promise for creating highly adaptive and tailored educational experiences in the future.

Healthcare

RL in healthcare enables the development of dynamic treatment regimes (DTRs) for chronic conditions, allowing providers to deliver tailored, adaptive interventions that improve patient outcomes. RL also enhances operational efficiency, optimizing resource allocation, scheduling, and workflow in hospitals while reducing costs. In drug discovery, RL accelerates the identification of effective compounds and predicts drug responses, saving time and resources.

RL excels in handling uncertainty and complexity, making it invaluable for predicting disease progression and selecting optimal treatments in critical care. By learning from data and outcomes, RL systems refine their strategies to enhance accuracy and reliability, supporting decision-making where traditional methods fall short.

Ordaōs utilizes reinforcement learning within its proprietary Design Engine to create de novo mini-proteins, known as miniPRO™, for therapeutic applications. This approach enables the rapid and efficient design of bespoke proteins that are more configurable, stable, and easier to manufacture than traditional antibodies, thereby accelerating the development of safer and more effective treatments.

UnityAI employs reinforcement learning (RL) to enhance hospital operations by optimizing patient flow and resource allocation. Their AI-driven platform analyzes real-time data to recommend actions that improve efficiency and patient care.

Despite its potential, RL faces challenges like data scarcity, as ethical concerns and privacy regulations limit access to real patient data, requiring simulated environments or historical datasets. Additionally, human physiology's complexity leads to partial observability, complicating decision-making. Effective application of RL depends on improving data availability, creating accurate simulations, and designing reward structures that balance short- and long-term health outcomes.

Marketing

In Marketing, RL can enhance strategies by enabling dynamic decision-making that adapts to consumer interactions in real-time. RL algorithms can personalize content delivery by analyzing individual customer behaviors and preferences, ensuring that marketing messages align closely with consumer interests.

Additionally, RL can optimize pricing strategies by continuously learning from market responses to different price points, thereby maximizing revenue. By leveraging RL, marketers can create more responsive and effective campaigns that evolve based on consumer feedback and behavior patterns.

For example, RL can facilitate automated A/B testing, with notable startups like Aampe and Offerfit using it to optimize marketing strategies dynamically.

Companies like Just Words and Motiva AI leverage RL to improve push notifications and email campaigns, respectively, while JewelML and Albatross apply it in product recommendation systems.

In the table below, we list startups together with their funding, investors, and a short description of what they do in RL.

When it comes to M&A activity in RL, we can note that most acquisitions do not list RL and name more broadly Artificial Intelligence as a reason for acquisition. That said, we managed to find some deals that occurred in recent years:

Remaining Challenges

Despite its huge potential, Reinforcement Learning is still nascent in its progress and potential. Below, we summarize some of the challenges that hinder RL from deeper penetration and impact:

• Data Scarcity and Quality: RL requires substantial amounts of high-quality data for effective training. In many fields, acquiring real-world data is still challenging due to ethical concerns, privacy regulations, skills, or cost.
• High Computational Costs: RL algorithms are computationally intensive, especially when simulating complex environments or processing high-dimensional data. Meta, for example, owns over 350,000 NVIDIA H100 GPUs for its computing power while most businesses might struggle to obtain a few.
• Reward Design Complexity: Designing appropriate reward structures that guide the agent effectively is difficult, particularly when balancing short-term and long-term goals. Part of the answer could be in leveraging hierarchical RL and multi-objective optimization to design reward mechanisms that align with complex, real-world objectives.
• Generalization Issues: RL models often struggle to generalize well to unseen environments or variations in real-world applications.
• Lack of Explainability: RL models can function as "black boxes," making it hard to explain their decisions, a significant concern in fields like healthcare and autonomous systems. This especially could be an issue in sensitive sectors like healthcare or defense. Developing methods to make RL algorithms interpretable, such as visualizing decision processes or integrating explainable AI (XAI) techniques could be part of the solution.
• Safety and Ethical Concerns: In critical applications like healthcare or autonomous driving, ensuring safety while the agent explores new strategies is a significant challenge. Implementing robust policies for ethical use, and addressing privacy, fairness, and safety concerns in sensitive applications can help alleviate some of these concerns.