PaperlayerView exact revision on arXiv

Source-linked AI summary

More than a Million Ways to Be Pushed: A High-Fidelity Experimental Dataset of Planar Pushing

Kuan-Ting Yu, Maria Bauza, Nima Fazeli, Alberto Rodriguez

arXiv:1604.04038v2cs.RO

TL;DR

Accurate prediction of pushed-object motion remains difficult because contact geometry and friction vary, while existing data-driven studies lack generality and common benchmarks. The paper builds an automated, high-fidelity planar-pushing dataset spanning six interaction dimensions and evaluates common frictional-pushing assumptions. The dataset contains more than a million timestamped pose-and-force samples for each shape-material combination, and the evaluation finds that some assumptions are reliable while friction coefficients vary across space, orientation, velocity, and time.

  • Problem

    Accurate pushing prediction is hindered by sensitivity to contact geometry and friction variability, while data-driven studies lack sufficient generality and common datasets or benchmarks.

  • Method

    The paper automates controlled planar-pushing experiments, recording pusher and object poses and interaction forces across six dimensions with force/torque sensing and Vicon tracking.

  • Results

    The dataset provides more than a million timestamped pose-and-force samples for each shape-material combination, and evaluates common analytical-model assumptions.

  • Takeaways & Limitations

    The study finds that the maximum power inequality generally represents friction-motion directions, whereas the coefficient of friction varies with space, orientation, velocity, and time.

Abstract

from arXiv · show

Pushing is a motion primitive useful to handle objects that are too large, too heavy, or too cluttered to be grasped. It is at the core of much of robotic manipulation, in particular when physical interaction is involved. It seems reasonable then to wish for robots to understand how pushed objects move. In reality, however, robots often rely on approximations which yield models that are computable, but also restricted and inaccurate. Just how close are those models? How reasonable are the assumptions they are based on? To help answer these questions, and to get a better experimental understanding of pushing, we present a comprehensive and high-fidelity dataset of planar pushing experiments. The dataset contains timestamped poses of a circular pusher and a pushed object, as well as forces at the interaction.We vary the push interaction in 6 dimensions: surface material, shape of the pushed object, contact position, pushing direction, pushing speed, and pushing acceleration. An industrial robot automates the data capturing along precisely controlled position-velocity-acceleration trajectories of the pusher, which give dense samples of positions and forces of uniform quality. We finish the paper by characterizing the variability of friction, and evaluating the most common assumptions and simplifications made by models of frictional pushing in robotics.

I. INTRODUCTION

Pushing is central to robotic manipulation and perception, yet accurate prediction is hindered by contact sensitivity, friction variability, limited data generality, and costly manual collection. The paper addresses these gaps with an automated, high-fidelity planar-pushing dataset spanning varied interactions.

  • Pushing supports positioning, reorientation, grasping under uncertainty or clutter, transportation, pose tracking, shape estimation, and inertial-parameter identification.These applications require models that predict how an object moves under a given push.
  • Small changes in contact geometry and friction variability hinder accurate prediction of pushed-object motion.
  • Data-driven pushing studies lack sufficient generality or treatment of variability, while common datasets and benchmarks remain scarce.This limits evaluation and comparison of learned models.
  • The study automates controlled pushing experiments and records high-sample-rate interactions using force/torque sensing and Vicon tracking.
  • The dataset varies surface material, object shape, contact position, pushing direction, speed, and acceleration, including dynamic pushing where inertia appreciably affects frictional forces.
  • The dataset provides more than a million timestamped pose-and-force data points for each combination of shape and material, alongside an evaluation of common analytical-model assumptions.

II. RELATED WORK

Prior pushing models use friction-based analytical representations and approximations to make motion prediction tractable. The related work includes quasistatic and dynamic formulations, uncertainty modeling, and approximations whose assumptions this paper evaluates.

  • Quasistatic interaction: Quasistatic interaction neglects inertia when object velocities are small, simplifying motion to a balance among contact, frictional, and gravity forces.This assumption is considered reasonable for many robotic-manipulation scales and speeds.
  • Analytical models: Foundational pushing models include voting theorems and limit-surface representations that map object motions to frictional forces.
  • Analytical models: Ellipsoidal limit-surface approximations reduce computational time and support closed-form quasistatic solutions for sticking and sliding.
  • Uncertainty and dynamics: Prior work addresses uncertainty by sampling pressure distributions, while dynamic studies assume either frictionless or infinitely frictional pusher-object interaction.The paper explicitly validates assumptions marked with an asterisk in the prior-work summary.

III. THE PUSHING DATASET

The dataset records planar pushing across six controlled dimensions, using high-fidelity pose and force measurements from automated experiments. Open-loop trajectories sampled at 250 Hz capture varied contact interactions, including sticking-to-sliding transitions.

  • III. THE PUSHING DATASET: The dataset varies object shape, surface material, contact position, push direction, speed, and acceleration.It includes 11 stainless-steel objects and four support surfaces, with multiple contact locations and push directions.
  • III. THE PUSHING DATASET: Push directions range from -80° to 80° around the contact normal in 20° increments, with 33–44 evenly spaced contact locations per object.
  • III. THE PUSHING DATASET: Pusher speeds span 10–500 mm/sec, while constant accelerations range from 0.1 to 2.5 ms−2, including zero acceleration for constant-speed trials.
  • III. THE PUSHING DATASET: Each experiment follows an open-loop trajectory in the object's initial frame, producing evolving contact geometry and capturing sticking-to-sliding transitions.Trajectories are executed and recorded at 250 Hz.
  • III. THE PUSHING DATASET: Position control provides accurate pusher position, speed, and acceleration while avoiding force-control challenges caused by friction constraints and coefficient errors.

IV. DATA COLLECTION SPECIFICATIONS

Data collection uses a six-degree-of-freedom industrial robot with a stiff cylindrical rod as the pusher. The setup is designed to execute controlled pushing motions.

  • IV. DATA COLLECTION SPECIFICATIONS: A 6 DOF industrial robotic manipulator equipped with a stiff cylindrical rod acts as the pusher.Figure 1 presents the experimental setup.

A. Hardware

The hardware combines an industrial robot, force-torque sensing, optical motion tracking, and standardized steel objects. Measurement accuracy is reported for both object and pusher poses.

  • Hardware: An ABB IRB 120 robot controls the tool center point's position, velocity, and acceleration with six degrees of freedom.Its 580 mm horizontal reach and 3 kg payload accommodate the approximately 1 kg objects.
  • Hardware: An ATI Gamma force-torque sensor measures pusher reaction forces with 1/160 N force resolution and 1/2000 N·m torque resolution.
  • Hardware: A Vicon system with five Bonita cameras tracks each object's pose using four reflective markers.Four asymmetric markers provide more stable readings than the three theoretically sufficient markers.
  • Hardware: Object position accuracy is below 0.5 mm for translation and 0.5° for rotation, while pusher pose accuracy is 0.1 mm.
  • Hardware: The stiff cylindrical steel pusher is 156 mm long and 9.5 mm in diameter, balancing occlusion reduction with rigidity.
  • Hardware: The dataset uses 11 bead-blasted stainless-steel objects, each 13 mm thick, with masses from 0.75 to 1.4 kg.The pusher-object friction coefficient is approximately 0.25.

B. Software

ROS integrates robot control, force-torque sensing, and motion tracking, while recording synchronized streams in reusable data formats.

  • B. Software: ROS publishes robot pose, object pose, and force-torque data as topics recorded at 250 Hz.Experiments are logged as ROS bag files and parsed into HDF5 and JSON formats.

C. Data Collection Process

The experiments use an automated loop that tracks the object, executes predefined straight pushes, records interaction data, and resets the object for iteration.

  • The tracker locates the object before the robot executes each open-loop straight push along a predefined position-velocity-acceleration trajectory.Vicon tracking and force-torque sensing record the interaction during execution.
  • A reset mechanism drags the object approximately to the plate center when needed, enabling repeated data collection.
  • Each push covers 5 cm and produces an average of 200 timestamped interactions involving pusher motion, object motion, and pushing force.The collection yields approximately 6,000 pushes per object and surface and more than a million interaction triples overall.

V. VARIABILITY OF DYNAMIC SURFACE FRICTION

The experiments characterize dynamic friction across spatial, temporal, speed, and directional variability using controlled scans and force measurements. Friction varies across materials and conditions, challenging a single constant-friction description.

  • The cage constrains rect1’s sliding while a wrist force-torque sensor measures the horizontal reaction force used to compute DCoF.The scan protocol varies location, repetition, speed, and direction.
  • Spatial DCoF variability differs by material, with standard deviations of 0.016 for delrin, 0.017 for abs, 0.024 for plywood, and 0.064 for pu.The distributions are measured over approximately 20 cm by 40 cm areas.
  • After 100 scans, DCoF decreases by 13.6% for abs, 22.2% for delrin, and 11.3% for plywood, while pu changes by -2.3%.The materials differ in break-in and abrasion behavior: pu shows little degradation over the tested force range.
  • DCoF varies little with speed for delrin, abs, and plywood, but increases up to 1.0 for pu at high speeds.Thus Coulomb friction is not a good approximation when experiments span a wide speed range.
  • Friction is close to isotropic for abs and delrin, slightly less so for plywood, and least isotropic for pu, whose largest-to-smallest friction ratio is around 3/2.

VI. EVALUATION OF MODELS OF FRICTIONAL SLIDING

The paper tests maximum-power behavior and ellipsoidal limit-surface models using controlled sliding twists and measured frictional wrenches. The simple models fit some materials better than others, especially poorly for pu.

  • The experiments vary translation-to-rotation velocity ratios, scan angles, starting distances, and object rotations to evaluate frictional sliding models.
  • With imposed object velocity, the maximum-power inequality is tested through ∆P = f · v − max_j(f_j · v) ≥ 0.The analysis uses data passing through a particular point to reduce effects from frictional variability.
  • All materials except pu produce ∆P values very close to 0; pu shows two regions with substantially negative ∆P near abrupt limit-curve transitions.
  • The measured limit surface resembles an ellipse but not exactly, with delrin most symmetric, abs and plywood slightly biased, and pu unlike an actual limit surface.The ellipsoidal fits assume a centered origin and estimate moment and force magnitudes from pure rotational and translational motion.

VII. STOCHASTICITY OF PUSHING MOTION

Repeated nominally identical pushes reveal structured, material-dependent uncertainty in final object pose. The resulting distributions are non-Gaussian, and experimental trajectories can differ markedly from deterministic simulator predictions.

  • The study repeats a particular straight-line push 2,000 times across four surface materials using fixed contact location, normal contact angle, 20 mm/s speed, and the same push setup.
  • Final pose distributions have at least three modes and are clearly not Gaussian.Displacements are represented as (∆x, ∆y, ∆θ) relative to an initial pose of (0, 0, 0).
  • The standard deviation of ending poses depends on the surface type, and the resulting spread is related qualitatively to friction variability.
  • The mean experimental trajectory and the model-driven simulator prediction look quite different.
  • Accurate vision and robot repetition do not eliminate appreciable, structured uncertainty in the outcome of a determined pushing interaction.The findings motivate further investigation of methods for representing frictional uncertainty or variability.

VIII. CONCLUSION

The paper presents a high-fidelity planar-pushing dataset spanning six experimental dimensions and more than a million timestamped measurements, then evaluates common modeling assumptions. It finds that directional friction assumptions are generally sound, while friction magnitude varies across space, orientation, velocity, and time.

  • Six dimensions—shape, surface material, contact location, pushing direction, velocity, and acceleration—span the planar-pushing experiments.
  • More than a million timestamped samples record pusher and slider positions together with interaction forces.
  • The dataset supports evaluation of common assumptions and approximations used in planar-pushing models.
  • The maximum power inequality generally represents the relationship between friction and motion directions.
  • The coefficient of friction is not necessarily constant, changing with space, orientation, velocity, and time.
  • Future work targets more accurate semi-parametric and stochastic friction models, simulation, planning, control, and out-of-plane motions such as rolling or toppling.
Loading 1604.04038v2…