Planning for Multi-stage Forceful Manipulation

Rachel M. Holladay, Tomás Lozano-Pérez, Alberto Rodriguez

Abstract

This paper introduces a framework for robots to plan complex, multi-step tasks that require significant force, such as opening a childproof bottle or cutting vegetables. By integrating "forceful kinematic chain" constraints into a task and motion planner, the system can reason about robot torque limits and grasp stability, ultimately selecting strategies that are robust to physical uncertainties like varying friction.

Introduction

Most robotic task and motion planning (TAMP) focuses on "pick-and-place" style movements where the primary constraints are geometric (avoiding collisions) and kinematic (reaching a pose). However, many everyday tasks—twisting a stubborn nut, slicing a cucumber, or opening a push-and-twist bottle—are "forceful." In these scenarios, the robot must exert a specific wrench (a combination of force and torque) on an object.

The childproof bottle task requires reasoning about the necessary wrench (force and torque) to open the cap while maintaining stability.

Problem Statement

Forceful manipulation introduces two major challenges:

  1. Stability (Fixturing): The object must be secured so it doesn't move while the robot pushes or twists it.
  2. Capability Limits: The robot’s own motors have torque limits, and its grip on a tool or object might slip if the required force is too high.

Furthermore, physical parameters like friction coefficients are rarely known perfectly. A plan that works on paper might fail in reality if the environment is slightly different than expected (e.g., a slippery table).

Methodology

The authors build upon PDDLStream, an existing TAMP framework, introducing several novel components:

  • Forceful Kinematic Chain Constraint: This model treats the entire system—from the robot's base through its joints to the contact point on the object—as a chain. The planner evaluates whether every "joint" in this chain (including frictional contacts) is stable under the required task wrench and gravity.
  • Limit Surfaces: To determine if a grasp will slip, the authors use mathematical models (ellipsoidal and polyhedral limit surfaces) to define the boundary of forces a contact patch can resist.
  • Cost-Sensitive Robust Planning: Instead of just finding any feasible plan, the authors formulate a "robust" version using Monte Carlo sampling to perturb physical parameters. They assign a cost to actions based on their probability of success ($Cost = -\log(P_{success})$).
  • Fixturing Reasoning: The planner can choose between various ways to secure an object: using a second robot arm, placing it in a vise, or using a high-friction surface.

Experimental Evaluation

The framework was evaluated on three custom-designed manipulation domains: the Childproof Bottle (requiring simultaneous push and twist), Nut Twisting (potentially using a spanner), and Vegetable Cutting (a two-stage press and slice). Experiments were conducted both in simulation and on a real-world ABB YuMi dual-arm robot.

Key Results and Insights

The system successfully navigated a wide range of strategies based on the specific constraints of the environment. A key finding was the trade-off between robustness and efficiency: as the demand for reliability increased, the planner shifted from "brittle" plans to robust ones, even if they required more steps (e.g., moving an object to a high-friction mat).

The framework evaluates multiple strategies in simulation, such as using a vise or a second arm, to find the most robust plan.

Other insights include:

  • Force as a Planning Constraint: Force must be considered during high-level planning, as it dictates necessary grasp poses and environmental interactions.
  • Lever Arm Optimization: In cutting tasks, the robust planner chooses to grasp the knife closer to the blade to reduce torque on the robot's hand, increasing stability.

Conclusion

This work demonstrates that reasoning about force and stability at the planning level allows robots to perform complex, forceful tasks reliably. While currently limited by a quasi-static assumption and the need for known models, future work integrating real-time tactile feedback promises even greater adaptability.

Summary updated

BibTeX
@inproceedings{holladay2021planning,
  author = {Rachel M. Holladay and Tomás Lozano-Pérez and Alberto Rodriguez},
  title = {{Planning for Multi-stage Forceful Manipulation}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2021},
  doi = {10.1109/icra48506.2021.9561233},
}

Loading related content...

Figure