SUN: Persistent Programs For Language-Grounded Control-to-Learning-to-Real Policies
This paper introduces SUN (Semantically UNified) Programs and the Kuafu system, which bridge model-based control and learned policies by automatically synthesizing typed, language-grounded executables that unify MPC verification and RL training, thereby enabling robust long-horizon manipulation without manual dense rewards or human demonstrations.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Robots have long been masters of the factory floor, repeating the same precise motion thousands of times without error. But when asked to perform a complex, multi-step task in a messy, unpredictable home—like opening a drawer, taking out a bottle, and placing it on a table—they often stumble. The difficulty lies in bridging the gap between two different ways of thinking about movement. One approach relies on rigid mathematical rules to calculate every joint angle and force, ensuring the robot does not crash, but this method is brittle and fails if the world changes slightly. The other approach uses trial-and-error learning, where a robot practices until it gets the hang of a task, but this often requires thousands of hours of human demonstration or carefully crafted rewards that are difficult to design. For years, these two methods have operated in separate silos, with the rigid planners unable to adapt and the learners unable to understand the deeper logic of the task.
Researchers at the University of California, Los Angeles, and their colleagues have introduced a new system called Kuafu that attempts to weave these two approaches together. Instead of treating the robot's instructions as a fleeting set of goals or a rigid script, they created a persistent, typed program that acts as a single source of truth for the entire process. This program, which they call a "Semantically Unified" or SUN Program, is written in a way that a computer can understand the physical relationships between objects, such as which way a drawer handle faces or how far a bottle must be lifted. Crucially, this same program is used to verify the robot's actions, to teach it how to move, and to generate the data needed for it to learn from scratch. By keeping the core meaning of the task constant from the initial planning stage all the way to the final learning phase, the system prevents the robot from drifting away from what it was actually supposed to do.
The system begins by taking a simple language instruction from a human, such as "open the drawer and place the cup inside." An artificial intelligence agent reads this instruction and constructs the SUN Program by selecting pre-defined building blocks that describe physical actions and constraints. It then tests this program in a high-fidelity simulation, using a sophisticated control method to see if the task is physically possible. If the simulation reveals that the instructions are flawed or impossible to execute, the system automatically repairs the program before any learning begins. This screening process is vital because it ensures that the robot is never asked to learn a task that cannot be done, filtering out bad ideas before they waste time and computing power.
Once the program is validated, the system uses it to generate thousands of successful examples of the robot performing the task. These examples are not just random movements; they are guided by the persistent program, which keeps track of every stage of the task, from the moment the gripper touches the handle to the moment the drawer is fully open. The robot then learns from these examples, first by mimicking the successful movements and then by refining them through a process of trial and error that is strictly bounded by the original program. This ensures that while the robot learns to be flexible and reactive, it never loses sight of the ultimate goal. The result is a policy that can execute complex, multi-stage tasks with a level of reliability that previous methods could not achieve.
In a series of tests involving nine different manipulation tasks, ranging from stacking cubes to reorienting bottles and opening drawers, the system demonstrated a significant improvement over existing methods. While other approaches that rely on sparse rewards or online planning struggled to succeed more than a third of the time, this new system achieved a success rate of over 82 percent. The researchers found that the system was particularly effective at handling tasks that required many steps, maintaining its performance even as the complexity of the sequence increased. Furthermore, the data generated by this system was of such high quality that it allowed a visual learning model to succeed in 46 percent of trials, a figure that was more than double the success rate of models trained on data from other generation methods.
The true test of any robotic system is whether it can move from the safety of a computer simulation to the real world. To verify this, the researchers deployed the trained policies on physical robots made by Franka and Kinova, using cameras to guide the robot's actions without any prior knowledge of the specific task structure. Without any additional tuning or real-world practice, the robots successfully completed 34.7 percent of the physical trials across various configurations. This zero-shot transfer, where a robot trained entirely in simulation works immediately on a real machine, suggests that the persistent program successfully captured the essential logic of the task, allowing the learned behavior to generalize to the physical world.
The researchers acknowledge that this approach has limits. It currently relies on a pre-defined library of physical actions and requires a clear understanding of the objects in the scene, meaning it cannot yet handle deformable objects like cloth or fluids without significant new programming. Additionally, the system advances through tasks in a single direction and cannot backtrack if a physical error occurs, which limits its ability to recover from certain types of mistakes. However, the results establish a new principle for robot learning: that keeping the semantic meaning of a task consistent across planning, verification, and learning creates a robust foundation for automation. By ensuring that the robot's understanding of the task does not drift, the system bridges the gap between rigid control and flexible learning, offering a path toward robots that can reliably perform complex tasks in the real world.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.