Smart Robots Learn to Dream: Breakthrough at Top Universities in Hong Kong and Beyond

Robots Learn to “Dream”: Top Universities Collaborate to Create Intelligent Robots with Imagination

When you dream at night, your brain simulates various possible scenarios and actions during sleep. This ability to “rehearse” helps us better navigate complex situations in reality. Now, research teams from leading institutions such as the University of Hong Kong, Tsinghua University, and others have successfully endowed robots with a similar “dreaming” capability. This groundbreaking research was published in February 2026, under the paper number arXiv:2602.11075v1, paving a new path for the development of robotic intelligence.

In the past, training a robot to master complex operations was as challenging as teaching a child to ride a bicycle. Traditional methods required robots to repeatedly attempt tasks in real-world environments, often falling and getting back up. This process was not only time-consuming and labor-intensive, but it also posed safety risks. Moreover, each practice session required manual resetting of the environment, resulting in exorbitant costs.

The research team discovered that even the most advanced robots struggled with tasks requiring precise operations, such as grabbing objects on a moving conveyor belt or handling soft, deformable materials. To address this challenge, they developed a revolutionary system called RISE, which stands for Reinforcement learning through Imagination for Self-Improvement. RISE allows robots to “dream” and practice in a virtual world. Similar to how human athletes mentally rehearse their actions before competitions, RISE enables robots to conduct extensive practice in imagined environments, subsequently translating these virtual experiences into real-world operational skills.

1. The Robot’s “Imagination Factory”: The Secret of the Combinatorial World Model

The core of the RISE system is an intelligent brain referred to as the Combinatorial World Model, acting as the robot’s dedicated dream factory. This factory consists of two precision workshops, each responsible for different tasks. The first workshop is the Dynamic Prediction Department, functioning like a skilled animator. When a robot intends to perform a specific action, this department quickly sketches potential future scenarios based on the current environmental conditions and the planned actions. For example, when a robot prepares to grab a moving block, the Dynamic Prediction Department simulates the block’s trajectory, the robotic arm’s movement path, and various possible outcomes of success or failure.

The efficiency of this prediction system is remarkable. Traditional simulation systems may take 10 minutes to generate 25 frames of multi-view predictions, while the Dynamic Prediction Department of RISE can accomplish the same task in less than 2 seconds, achieving a 300-fold increase in efficiency. This astonishing speed allows robots to engage in extensive virtual practice, akin to an athlete performing thousands of mental rehearsals in a short time.

The second workshop is known as the Value Assessment Department, which functions like an experienced coach. After the Dynamic Prediction Department generates various potential future scenarios, the Value Assessment Department meticulously analyzes each scene to determine which actions are favorable and which are not. This department provides not just binary success or failure judgments but detailed scoring evaluations, similar to judges scoring a gymnastics competition. The training process for the Value Assessment Department is particularly intriguing. The research team trained it using two different evaluation methods: Progress Estimation, which gradually increases scores over time, and Temporal Difference Learning, a more sensitive method capable of capturing minute errors during operations. By combining these two approaches, the Value Assessment Department achieves both stable scoring and the sharp detection of issues.

Notably, these two workshops do not operate independently; they work in close collaboration. The Dynamic Prediction Department generates imagined scenarios, the Value Assessment Department evaluates them, and then feedback is provided to the robot’s decision-making system. This collaboration forms a complete learning cycle: imagine, assess, improve, and then repeat, continuously enhancing the robot’s skills.

2. From “Novice” to “Expert”: The Robot’s Learning Progression

The training process of the RISE system resembles cultivating a skill master from an amateur level, divided into two key stages, each with unique learning methods and objectives. The first stage, termed Policy Warm-Up, serves as a foundational training camp for the robot. During this stage, the robot learns the most basic operational skills, similar to how a new driver familiarizes themselves with the steering wheel and brakes. The research team provides the robot with a wealth of real operational videos, including expert demonstrations, successful cases, and even instances of manual error correction. During this phase, the robot acquires a special skill: adjusting its behavior based on advantage signals. This is akin to equipping the robot with an internal voice that indicates “this action is good, keep it up” or “this may lead to issues, proceed with caution.” When the robot receives a high advantage signal, it executes actions with greater confidence; conversely, a low advantage signal prompts it to be more cautious or adopt alternative strategies.

The second stage is where the RISE system truly shines, known as the Self-Improvement Loop. In this phase, the robot begins extensive “dream training” in the virtual world. The entire process resembles an unending learning cycle, consisting of two alternating steps. In the Virtual Practice step, the robot starts from a real environmental state and then imagines various actions. The system inputs optimized behavior commands into the robot’s virtual avatar, enabling it to perform at its best in the dream environment. Subsequently, the Combinatorial World Model generates potential future scenarios from these actions and provides detailed scoring for each scenario. This process allows the robot to mentally rehearse thousands of times, receiving comprehensive feedback after each practice. To ensure diversity in training, the system also utilizes these imagined scenarios as new starting points for further in-depth virtual practice. However, the research team discovered that more than two consecutive rounds of virtual practice could lead to cumulative prediction errors affecting the training outcomes, akin to how information in a game of telephone can become distorted after multiple transmissions.

During the Skill Upgrade step, the robot analyzes all its virtual experiences to learn how to perform tasks better in the real world. The system designates high-scoring actions from virtual practice as learning targets, allowing the robot to make better choices when faced with similar situations. Moreover, to prevent the robot from forgetting previously learned foundational skills, the system regularly reviews real-world operational experiences. The advantages of this learning approach are clear. Traditional robot training requires thousands of attempts in real environments, where each failure could lead to equipment damage or safety risks. In contrast, the RISE system enables robots to engage in extensive practice in a secure virtual environment, avoiding real-world risks and significantly enhancing learning efficiency.

3. Real-World Testing: Outstanding Performance in Three Challenging Tasks

To validate the practical effectiveness of the RISE system, the research team designed three highly challenging real-world tasks, each testing different aspects of the robot’s capabilities. These tasks presented enough difficulty to deter traditional robotic systems, yet the RISE system demonstrated remarkable performance improvements across all tasks. The first task, Dynamic Block Sorting, resembles a challenging grabbing game for robots on a rapidly moving conveyor belt. Blocks move at different speeds, and the robot must accurately identify the color of each block and place them in corresponding color storage boxes. This task not only tests the robot’s visual recognition ability but also critically assesses its tracking and grabbing precision of dynamic targets. In this task, the RISE system achieved an 85% success rate, a significant boost from the 35% of the base system, representing a 50 percentage point increase.

The second task, Backpack Packing, simulates the everyday process of organizing luggage. The robot must open a soft backpack, stuff clothing inside, then lift the backpack to allow the clothes to settle at the bottom, finally zipping it up to complete the packing. This task particularly tests the robot’s capability to handle soft, deformable objects, as both the backpack and clothes are not rigid and their shapes change during the operation. The RISE system excelled in this task as well, with the success rate soaring from traditional methods’ 30% to 85%, marking a remarkable increase of 55 percentage points. The third task, Box Packaging, involves a complex operation that requires precise coordination of both robotic arms. The robot must place a cup into a box, sequentially fold the side flaps and the back flap, and finally insert a lock correctly into a slot to complete the packaging. Any deviation in this process could lead to failure. In this highly challenging task, the RISE system achieved an astonishing 95% success rate, compared to the 35% of the base method, representing a massive improvement of 60 percentage points.

The implications of these experimental results extend beyond mere numbers. The research team found that the RISE system not only significantly improved success rates, but also exhibited enhanced adaptability and robustness. Traditional robotic systems are prone to failure with even minor environmental changes or unexpected situations, whereas the robots trained with the RISE system can better handle these uncertainties. To ensure fairness in experiments, the research team also compared the RISE system with various advanced benchmark methods, including traditional imitation learning, online reinforcement learning, and other cutting-edge robotic training techniques. In all comparisons, the RISE system demonstrated a clear advantage, particularly in tasks requiring precise operations and dynamic adaptability.

4. In-Depth Analysis: Key Factors Behind RISE System’s Success

The exceptional performance of the RISE system can be attributed to several meticulously designed key factors, each validated and optimized through extensive experimentation. First is the implementation of the Task-Centric Batch Processing strategy. While training the Combinatorial World Model, the research team identified a critical issue: mixing data from different tasks and scenarios during training significantly reduces the model’s learning effectiveness. This is akin to having a student simultaneously learn math, language, and history, leading to distracted attention and decreased learning efficiency. To address this, the research team adopted a clever training strategy. They ensured that each training batch focused on data from the same type of task while incorporating various action combinations under that task. This approach is similar to having a student concentrate on learning math for a specific period while encountering a variety of math problems, thus maintaining focus while ensuring a comprehensive understanding. Experiments demonstrated that this strategy not only improved the model’s convergence speed but also significantly enhanced the operational capabilities of the robots produced.

Secondly, precise control of the Offline Data Mixing Ratio plays a crucial role in the robot’s learning process, balancing two types of experience: offline data collected from the real world and online data generated in virtual environments. The research team found that the mixing ratio of these two data types critically influences the final outcomes. Through extensive experiments, they determined that the optimal mixing ratio is 60% offline data to 40% online data. When the proportion of offline data is too low (for instance, only 10%), the robot tends to forget foundational operational skills, resulting in a significant decrease in success rates for simple tasks. Conversely, when the proportion of offline data is too high (reaching 90%), the robot becomes overly conservative and fails to fully leverage new skills learned from virtual practices. The 60:40 golden ratio ensures that the robot maintains stable foundational abilities while continuing to learn and improve.

The third key factor is the design of the Advantage Conditioning mechanism. This mechanism enables robots to adjust their behavioral strategies based on the current levels of advantage. The research team categorized possible advantage values into ten levels, essentially providing the robot with a “confidence index” from 1 to 10. When the robot receives a high-level advantage signal, it proactively executes complex operations; when it receives a low-level signal, it opts for more conservative and cautious strategies. The brilliance of this design lies in its simulation of human expert decision-making processes. Experienced operators choose different operational strategies based on the complexity of current tasks and their confidence levels.

The final key factor is the Dual Value Learning method. Traditional robotic training typically employs a single evaluation method, focusing either on overall task progress or on the success/failure of specific steps. The RISE system innovatively combines two complementary learning methods: Progress Estimation and Temporal Difference Learning. The Progress Estimation method enables the robot to understand the overall context and timing of tasks, akin to providing it with an internal timetable. Meanwhile, Temporal Difference Learning makes the robot sensitive to subtle variations in the operational process, allowing it to promptly detect and correct minor deviations. The combination of these two methods ensures that the robot maintains a clear understanding of its larger goals while making precise adjustments during execution.

5. Breaking Boundaries: Technical Innovations and Limitations of the RISE System

While the RISE system has achieved several significant technical breakthroughs, the research team candidly acknowledges some current limitations of the system, highlighting areas for future improvements. In terms of technical innovation, the RISE system’s most significant breakthrough lies in successfully integrating imaginative capabilities into robotic learning. Traditional robot training relies on repeated trial and error in real environments, a method that is not only costly but also carries safety risks. The RISE system constructs high-quality virtual environments, allowing robots to practice extensively in a safe imaginative space, akin to providing them with a dedicated training simulator. The effects of this imaginative training are substantial. In the dynamic block sorting task, a robot equipped with the RISE system effectively completed tens of thousands of grabbing practices in the virtual environment, a volume of practice that would require months and incur significant costs in real-world settings. Through virtual training, the entire process can be completed in just a few days.

Another important innovation is the design concept of the Hierarchical Architecture. The RISE system decomposes complex robotic learning problems into two relatively independent yet closely cooperating modules: the dynamic model responsible for predicting the future and the value model responsible for evaluating actions. This decomposition not only makes the system easier to understand and debug but, more importantly, allows each module to adopt the most suitable technological solutions. The dynamic model employs advanced video generation techniques to swiftly produce high-quality future scenario predictions, while the value model is based on large-scale language model architectures, inheriting advantages in complex reasoning. This “division of labor” design ensures the efficient operation of the entire system.

However, the RISE system also faces limitations that cannot currently be fully overcome. The most significant limitation arises from the gap between the virtual world and the real world. Although the world model of RISE is quite advanced, the virtual scenes it generates cannot completely replicate the complexities of the real world. This is particularly true when dealing with rare or extreme conditions, where the effectiveness of virtual training may be compromised. The research team found that when faced with scenarios less frequently encountered in training data, the world model occasionally generates physically unrealistic predictions. For instance, when simulating the deformation of soft objects, the model may sometimes produce results that violate the laws of physics. While such occurrences are infrequent, they do limit the system’s reliability in certain extreme situations.

Another limitation is the demand for computational resources. Although the RISE system significantly reduces costs compared to direct real-world training, it still requires substantial computational power. Training a complete RISE system necessitates the use of multiple high-performance GPUs for several days to a week, which may pose a challenge for research institutions or companies with limited resources. Additionally, there is room for improvement when the RISE system addresses tasks requiring long-sequence reasoning. While the system has excelled in current test tasks, these tasks have relatively short time spans. Further validation is needed to assess how the system performs with complex tasks that take several minutes or longer to complete. Lastly, the RISE system is currently optimized primarily for operational tasks; its applicability to tasks requiring complex language understanding or multimodal reasoning still needs to be verified. Although the system’s foundational architecture supports these extensions, the specific implementations and optimizations will require substantial additional research efforts.

Despite these limitations, the research team remains confident about the future of the RISE system. They believe that with continuous advancements in computational technology and ongoing improvements in the accuracy of world models, these current limitations will gradually be addressed. More importantly, the RISE system has already demonstrated the feasibility of training robots through imagination, opening a new avenue for development in the entire field. Ultimately, the RISE system represents a significant breakthrough in the realm of robotic learning. It not only achieves multiple innovations on a technical level but also provides a novel possibility for robots to attain human-like intelligence. By enabling robots to learn to “dream” and “imagine,” we are witnessing artificial intelligence advancing toward a more intelligent and autonomous direction. The significance of this research lies not only in the achievements accomplished but also in its guidance for the future development of robotic technology. As technology continues to evolve, we have every reason to believe that intelligent robots with imaginative capabilities will soon become indispensable partners and assistants in our lives.

Q&A

Q1: How does the RISE system enable robots to learn to imagine?
A: The RISE system equips robots with imaginative capabilities by constructing a Combinatorial World Model. This model consists of two core components: the Dynamic Prediction Department generates potential future scenarios based on the current environment and planned actions, similar to an animator sketching future images; the Value Assessment Department evaluates these imagined scenarios. Robots learn and improve skills by repeatedly practicing in these virtual scenarios.

Q2: How much has the training efficiency improved compared to traditional methods?
A: The RISE system has significantly enhanced training efficiency. In terms of generating predictions, RISE completes in 2 seconds what traditional systems take 10 minutes to accomplish, achieving a 300-fold increase in efficiency. In actual task performance, RISE achieved success rates of 85%, 85%, and 95% in dynamic block sorting, backpack packing, and box packaging tasks, respectively, compared to base methods, which improved by 50, 55, and 60 percentage points.

Q3: What complex tasks can robots trained with the RISE system handle?
A: Robots trained with the RISE system can handle various challenging operational tasks, including accurately grabbing and sorting colored blocks on a moving conveyor belt, packing soft, deformable backpacks and clothing, and performing box packaging tasks that require precise coordination of both robotic arms. These tasks demand dynamic adaptability, precise operations, and complex reasoning capabilities, far exceeding the capabilities of traditional robots.