Abstract
Multi-stage forceful manipulation tasks, such as twisting a nut on a bolt, require reasoning over interlocking constraints over discrete and continuous choices. The robot must choose a sequence of discrete actions, or strategy, such as whether to pick up an object, and the continuous parameters of each of those actions, such as how to grasp that object. In forceful manipulation tasks, the force requirements substantially impact the choices of both strategy and parameters. To enable planning and executing forceful manipulation, we augment an existing task and motion planner with controllers that exert wrenches and constraints that explicitly consider torque and frictional limits. In two domains, opening a childproof bottle and twisting a nut, we demonstrate how the system considers a combinatorial number of strategies and how choosing actions that are robust to parameter variations impacts the choice of strategy.
Introduction
Most robotic planning frameworks focus primarily on kinematics—determining collision-free paths from a start to a goal configuration. However, many essential real-world tasks, such as unscrewing a stuck bolt or opening a childproof container, are fundamentally "forceful." In these scenarios, the forces applied to manipulate an object also act upon the environment and the robot itself. If the object is not properly secured, the applied force will cause it to slip or slide rather than achieve the desired manipulation.
Forceful manipulation introduces a combinatorial explosion of choices. A robot must decide on a high-level strategy (e.g., using a palm vs. fingers, or employing a vise) while simultaneously optimizing continuous parameters (e.g., grasp location, downward pressure). Traditional geometric planners fail in these domains because they do not account for physical impossibilities arising from torque or friction limits.
Methodology: Force-Aware TAMP
The proposed framework extends Task and Motion Planning (TAMP) by integrating force-based constraints and friction models. Built upon the PDDLStream framework, the system incorporates several key innovations:
1. Force-Based Constraints
The planner treats "wrenches" (combinations of force and torque) as state variables. It ensures that every link in the kinematic chain—from the robot's actuators to the contact points between objects and surfaces—can withstand the required forces without exceeding physical limits.
2. Stability and Friction Modeling
To predict and prevent slippage, the system employs mathematical models of contact mechanics. This includes ellipsoidal limit surfaces for point contacts (like fingertips) and polyhedral friction cones for larger planar contacts (like a bottle base on a table).
3. Robust Planning and Cost Sensitivity
The planner implements cost-sensitive reasoning to favor strategies that are robust to environmental uncertainties. By simulating small perturbations in parameters like friction coefficients, the system calculates a "robustness cost," steering the robot toward strategies that remain viable even if real-world conditions deviate slightly from the model.
4. Cartesian Impedance Control
Execution is handled by a controller that models the robot arm as a mass-spring-damper system. This allows the robot to exert precise wrenches while maintaining the compliance necessary to handle contact-rich interactions safely.
Experimental Evaluation
The system was evaluated in two challenging domains: opening a childproof bottle and twisting a nut.
Childproof Bottle Opening
This task requires a simultaneous downward force and a twisting torque. The robot can choose between various "twisters" (fingers, palm, or a specialized tool) and "fixturing" methods (using a second robot arm, a vise, or high-friction mats). The planner successfully navigated 16 different strategy combinations.
Nut Twisting
In this domain, the robot must use a spanner or its fingers to turn a nut. It must also ensure the stability of the beam holding the bolt, potentially by placing a heavy weight on it. The planner demonstrated the ability to balance the trade-off between stability (heavier weights) and placement accuracy (lighter weights).
Key Insights and Results
- Force-Driven Strategy: Force requirements are the primary drivers of high-level planning. If a surface is too slippery, the planner automatically pivots from a simple push-and-twist strategy to a more complex one involving a vise or a second arm.
- Constraint Propagation: The system treats the entire setup—robot, tools, and environment—as a single kinematic chain, ensuring stability from the fingertip to the robot's base.
- Efficiency: Despite the complexity of the search space, the planner typically finds solutions within 25 to 180 seconds depending on the strategy complexity.
Discussion and Future Work
While the framework successfully integrates force into TAMP, several areas for improvement remain. The current system uses fixed compliance levels; future iterations could plan for variable stiffness. Additionally, the planning time remains high for real-time applications, and the execution is currently open-loop, meaning the robot cannot yet adapt to unexpected resistance without re-planning. Finally, the current model assumes quasi-static physics, limiting its application to slow-moving tasks rather than high-speed dynamic actions like hammering.
Summary updated