Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Precise Object Placement Using Force-Torque Feedback

Osher Lerner, Zachary Tam, Michael Equi

TL;DR

This work tackles precise object placement in challenging environments by using force-torque feedback to recover from planning, execution, and sensor noise errors. It presents a pipeline that leverages a wrist-mounted force-torque sensor to reconstruct contact information and guide iterative pose refinements along the surface tangent, aided by a quasi-static physics model and surface normal optimization. On 46 stacking trials, the method achieves $100\%$ success when no adjustment is needed and $17\%$ success when adjustments are required, demonstrating robustness to noise and kinematic inaccuracies while exposing limitations on non-horizontal surfaces and near-stable poses. The study highlights the practical impact of FT-based placement for robust stacking and points to future directions, including small-offset detection, handling irregular objects, and integrating learning-based strategies to improve convergence and generalization.

Abstract

Precise object manipulation and placement is a common problem for household robots, surgery robots, and robots working on in-situ construction. Prior work using computer vision, depth sensors, and reinforcement learning lacks the ability to reactively recover from planning errors, execution errors, or sensor noise. This work introduces a method that uses force-torque sensing to robustly place objects in stable poses, even in adversarial environments. On 46 trials, our method finds success rates of 100% for basic stacking, and 17% for cases requiring adjustment.

Precise Object Placement Using Force-Torque Feedback

TL;DR

This work tackles precise object placement in challenging environments by using force-torque feedback to recover from planning, execution, and sensor noise errors. It presents a pipeline that leverages a wrist-mounted force-torque sensor to reconstruct contact information and guide iterative pose refinements along the surface tangent, aided by a quasi-static physics model and surface normal optimization. On 46 stacking trials, the method achieves success when no adjustment is needed and success when adjustments are required, demonstrating robustness to noise and kinematic inaccuracies while exposing limitations on non-horizontal surfaces and near-stable poses. The study highlights the practical impact of FT-based placement for robust stacking and points to future directions, including small-offset detection, handling irregular objects, and integrating learning-based strategies to improve convergence and generalization.

Abstract

Precise object manipulation and placement is a common problem for household robots, surgery robots, and robots working on in-situ construction. Prior work using computer vision, depth sensors, and reinforcement learning lacks the ability to reactively recover from planning errors, execution errors, or sensor noise. This work introduces a method that uses force-torque sensing to robustly place objects in stable poses, even in adversarial environments. On 46 trials, our method finds success rates of 100% for basic stacking, and 17% for cases requiring adjustment.
Paper Structure (17 sections, 4 equations, 5 figures)

This paper contains 17 sections, 4 equations, 5 figures.

Table of Contents

  1. INTRODUCTION
  2. RELATED WORK
  3. Force-Torque Feedback
  4. Stone Stacking
  5. Alternative Stacking
  6. METHODS
  7. Pipeline
  8. Physics Model
  9. Surface Normal Optimization
  10. Calibration
  11. RESULTS
  12. Physical Setup
  13. Stacking Experiments
  14. Algorithm Component Experiments and Analysis
  15. Implications
  16. ...and 2 more sections

Figures (5)

  • Figure 1: Left: The force diagram of a grasp before placement (normal forces at the jaws omitted for clarity). Right: The new forces on the rock during a place attempt, and the resulting torque.
  • Figure 2: Left: The simplified force diagram of an attempted rock placement, used to compute the normal force at rest. Middle: The torque induced on the rock COM, which we reverse engineer to find the contact point r from a measured torque. Right: The normal force is along the surface normal of the tower. Since the tower is likely convex around the point of contact, this can be used to find the direction towards a flat peak.
  • Figure 3: Norms of force and torque readings upon contact during vertical descent. Left: Centered placement. Right: Off-center placement.
  • Figure 4: A six-object stack constructed from the ground-up by our policy.
  • Figure 5: Left: Experimental readings of characteristic states from 3 place maneuvers. Upon first contact, noise still overwhelms torque readings. After pressing with 10N, the direction of the stable site can be inferred from the data. Right: Plotted center of mass for each candidate pose and the algorithm's suggested shift. "Centered" in green, "Y-shift" in red, "X-shift" in blue, and true tower center in black.