- The paper introduces a reactive TAMP method that integrates an Active Inference planner with a Multi-Modal MPPI controller for adaptive decision-making.
- The approach converts continuous world states into symbolic observations using behavior trees to generate and evaluate multiple alternative plans.
- Real-world experiments demonstrate the method’s superiority over single-strategy approaches and a reinforcement learning baseline under disturbances.
Introduction to Reactive Task and Motion Planning
Task and Motion Planning (TAMP), a significant area in robotics, involves creating plans that incorporate both discrete actions and continuous movements. Traditional approaches often assume a static environment and generate plans that are executed without considering changes that might occur in the real world. The dynamic nature of real-world scenarios, such as moving obstacles or unexpected disturbances, necessitates a more reactive planning approach that can adapt both high-level actions and low-level motions in real-time.
Active Inference Planner and Model Predictive Path Integral Controller
To address this need, an integrated planning strategy is introduced, combining two key components: an enhanced Active Inference planner (AIP) and a novel Multi-Modal Model Predictive Path Integral controller (M3P2I). The Active Inference planner is designed for high-level decision-making and can generate multiple alternative plans based on the current state and available actions. Each plan is linked to a different cost function, which is then evaluated by the M3P2I.
The M3P2I is based on a sampling approach that can handle complex dynamics and non-linearities, often characteristic of manipulation tasks. It extends this approach by enabling the parallel sampling of multiple plan alternatives, thus considering different possible actions at the same time. The AIP and M3P2I work together to create a robust control scheme that can handle uncertainties and adapt fluidly at both the action and motion level.
Algorithm and Implementation Details
The method operates by first converting continuous world states into symbolic observations, which are then processed by behavior trees and Active Inference to determine desired states or actions. The AIP then generates alternative action plans considering these desired states. Each plan is linked to specific cost functions, which are forwarded to the M3P2I for evaluation. The M3P2I samples control input sequences across these alternatives using a physics simulator, leading to multiple trajectory rollouts. By assigning weights to these rollouts based on their cost, the controller computes an optimal control input that effectively blends various strategic approaches.
Evaluation and Results
The proposed reactive TAMP method was tested in scenarios involving a push-and-place task and an object stacking task with a robotic manipulator. The results showed that combining the push and pull strategies using the M3P2I led to successful plan execution in cases where single-strategy approaches failed. Furthermore, the method performed favorably compared to a reinforcement learning baseline when faced with disturbances, despite requiring no prior training. Real-world experiments also demonstrated the approach's effectiveness in dealing with dynamic obstacles and disturbances.
In summary, by leveraging the combined strengths of the Active Inference planner and the Multi-Modal Model Predictive Path Integral controller within a reactive TAMP framework, this work presents a significant step towards enabling robots to perform complex tasks in unpredictable and dynamic environments.