Abstract
Task and motion planning (TAMP) algorithms help robots plan behaviors in discrete and continuous spaces. However, real-world scenarios often involve unknown object properties (e.g., size and weight), making traditional planning difficult. This paper introduces Task-Motion Object-Centric planning (TMOC), a grounded TAMP algorithm that learns to ground objects and their physical properties using a physics engine. TMOC is particularly effective for dynamic, multi-object interactions. Evaluated in simulation and on a real UR5e robot, TMOC outperforms competitive baselines in cumulative utility.
Introduction: The Challenge of Unknown Physical Properties
Traditional TAMP algorithms typically assume a perfect "world model," where the robot knows exact dimensions, weights, and friction coefficients of all objects. In practice, such as in a kitchen or warehouse, these properties are often unknown. A robot might fail to stack blocks if it doesn't realize the bottom block is too light to support the others. This paper addresses the problem of planning when object properties, state mappings, and action transition functions are unknown.
Methods: Task-Motion Object-Centric Planning (TMOC)
The authors propose the Planning with Physics-based Simulation (PPS) framework and the TMOC algorithm. PPS integrates a high-fidelity physics engine to bridge symbolic states (e.g., is_holding(block)) with continuous physical parameters.
The TMOC algorithm operates through several key mechanisms:
- Optimistic Initialization: Following the R-max principle, the robot initially assumes actions will succeed.
- Particle-based Simulation: The robot maintains $N$ "particles," representing different hypothesized object properties.
- Dual-Loop Learning: Real-world actions are mirrored in $N$ simulated worlds. By comparing outcomes, the robot updates particle weights (similar to a particle filter), effectively learning true physical properties over time.
- Integrated Planning: Learned properties refine both high-level task strategies and low-level motion trajectories.
Evaluation and Results
The framework was tested in a "Stack-and-Push" domain using both Pybox2D simulation and a real UR5e robotic arm.
Key Results
- Superior Performance: TMOC significantly outperformed baselines such as TMP-RL, GP, and TOEP in cumulative utility. As the robot grounded object properties, its utility curve rose sharply.
- Convergence and Efficiency: In simulation, TMOC achieved near 100% optimal plans after 5,000 episodes. Crucially, real-world learning was efficient, with the UR5e robot learning reliable mappings within 32 to 44 episodes (approx. 12 hours).
Discussion and Future Work
TMOC demonstrates that "grounding"—mapping symbols to physical reality—is a learning problem solvable by treating simulation parameters as latent variables. While successful in 2D-constrained environments, future work aims to extend this to unconstrained 3D spaces and optimize the computational cost of running parallel simulations. Additionally, learning to identify which objects in a scene are task-relevant could further improve efficiency.
Summary updated