TLDR: RAE+UPOM is a novel integrated actor-planner system that uses shared hierarchical operational models for both planning and acting on mobile robots. It addresses the common challenge of inconsistency between abstract planner models and real-world robot control. Deployed on a Mobipick robot for object collection, the system demonstrated robust task execution under action failures and sensor noise, intelligently adapting and recovering from issues like navigation or perception failures. It shows how detailed operational models and Monte Carlo planning can enable robots to make effective, real-time decisions in uncertain environments.
Robotics has made incredible strides, but a persistent challenge remains: bridging the gap between abstract plans generated by AI and the complex, unpredictable reality of a robot operating in the real world. Traditional robotic systems often rely on two separate models: a “descriptive model” for planning, which simplifies how actions work, and an “operational model” for execution, which handles the nitty-gritty details of real-time control and unexpected events. This disconnect can lead to plans that look good on paper but fail when faced with the messiness of reality.
A groundbreaking new research paper, “Acting and Planning with Hierarchical Operational Models on a Mobile Robot: A Study with RAE+UPOM,” introduces an integrated actor-planner system designed to overcome this very challenge. This system, called RAE+UPOM, is unique because it uses a single set of “hierarchical operational models” for both planning and acting, ensuring that the robot’s brain and its physical actions are always on the same page. You can find the full research paper here: Acting and Planning with Hierarchical Operational Models on a Mobile Robot: A Study with RAE+UPOM.
What is RAE+UPOM?
RAE stands for Reactive Acting Engine, and UPOM is an anytime UCT-like Monte Carlo planner. In simple terms, RAE is the part that executes actions and reacts to the environment, while UPOM is the intelligent planner that helps RAE decide the best course of action. They work together in an interleaved fashion, meaning planning and acting happen almost simultaneously, constantly adapting to new information and unexpected situations.
The core idea is that instead of having a simplified model for planning and a complex one for execution, RAE+UPOM uses rich, detailed operational models for both. These models incorporate complex control structures, error handling, and real-time decision-making, making the planning process much more realistic and robust. When a task needs to be performed, UPOM simulates various ways (called “refinement methods”) to achieve it, considering probabilistic outcomes and costs, and then recommends the best method to RAE. RAE then executes this method, and if something goes wrong, it can retry or explore alternative methods suggested by UPOM.
Real-World Deployment: The Mobipick Robot
The researchers put RAE+UPOM to the test on a real physical robot called Mobipick, a mobile manipulator equipped with a UR5 arm and a camera. The task was an “object collection task” in a hardware lab environment. The robot had to identify various objects scattered across multiple tables, pick them up, and place them on a designated target table. Crucially, the robot had no prior knowledge of the objects’ exact locations and had to contend with action failures, sensor noise, and uncertainties in the real world.
The hierarchical operational models allowed the robot to handle complex sub-tasks like exploring tables, perceiving objects, driving, and manipulating items. For instance, perception was modeled in two steps: a coarse initial scan to find objects, followed by a fine-grained perception step before grasping to accurately estimate an object’s pose. This detailed modeling, shared between the actor and planner, enabled the robot to adapt its camera angles and movements to ensure successful object detection and collection.
Key Findings and Robustness
The experiments demonstrated impressive robustness. RAE+UPOM successfully executed tasks even when faced with common robotic challenges like cable entanglement (leading to an emergency stop), navigation failures (where the robot couldn’t reach a target table), and perception failures (where an object wasn’t recognized). In these scenarios, the system didn’t just give up; it intelligently retried actions or selected alternative strategies, showcasing its ability to recover and complete tasks under adversity.
One interesting finding was the impact of using a “transport box.” While the box itself had no utility value, the system learned through planning that using it to transport multiple objects simultaneously was more efficient, leading to faster task completion and higher overall utility. This highlights how the system’s utility function, which favors early collection of high-value objects, combined with its ability to simulate different strategies, leads to emergent intelligent behaviors.
The research also provided insights into the planner’s internal decision-making. By visualizing the “rollouts” (simulated executions) performed by UPOM, the researchers could see how the planner explored different strategies and adapted as it gathered more information about the environment. Initially, many rollouts might yield no utility as the robot explores, but as it perceives objects, the success rate of its simulated plans increases.
Also Read:
- Neurosymbolic AI: Enabling Smarter Robots Through Combined Perception and Knowledge
- Robots Navigate Smarter with Cognitive Demand-Driven System
Future Implications
This work represents a significant step forward in integrated acting and planning for robotics. By removing the traditional gap between planning models and real-world implementations, RAE+UPOM enhances reactivity, maintainability, reusability, and robustness in robotic systems. The researchers plan to further investigate how the system handles dead ends, more dynamic events, and different failure types, as well as exploring the use of machine learning to improve refinement method selection and prioritize rollouts.
This integrated approach promises a future where robots can operate more autonomously and reliably in complex, unpredictable environments, adapting to challenges on the fly without constant human intervention.
