Learning an object by nudging it
A robot preparing to move an unfamiliar object needs more than a picture of its exterior. Mass determines the force needed to accelerate it, while the location of its center of mass determines whether a push, lift or grasp will remain stable. A monitor with most of its weight high above the table behaves differently from a uniform box, even if their visible outlines appear similarly manageable.
Researchers at Worcester Polytechnic Institute have tested a way to estimate an object’s mass and center-of-mass height without first grasping it or deliberately knocking it over. Their first-version preprint, submitted on September 11, describes an ABB IRB120 manipulator that slowly pushes and retracts against an object while measuring force and observing the resulting tilt. The method then fits those measurements to a mechanical model. It is a laboratory result involving four objects, not evidence that robots can safely probe arbitrary possessions or production equipment.
Force reveals what appearance cannot
The center of mass is the point at which an object’s distributed weight can be treated as acting for many mechanical calculations. Its horizontal location affects where the object balances on its support. Its height influences how quickly the object approaches instability when pushed from above. That vertical component is difficult to infer from an ordinary camera when the internal distribution of material is unknown.
The proposed procedure starts with an estimate of the center of mass projected onto the supporting surface. That planar estimate helps select a high pushing location intended to create useful tipping leverage without making the object slide. A rigid, 3D-printed fingertip then applies a slow horizontal push at a constant height. A six-axis force-and-torque sensor at the robot’s wrist records the interaction, while a fixed camera observes an AprilTag attached to the object to recover its changing angle.
As the object rotates around an edge, the force needed to hold it changes in a way governed by its weight, geometry and center-of-mass position. The researchers collect a coupled force-angle trajectory during both the outward push and the return motion. They filter measurement noise, use a linear fit to obtain an initial estimate, reject statistical outliers and refine the parameters with nonlinear least-squares optimization. The resulting model estimates mass, center-of-mass height and the angle at which toppling would begin.
The return motion helps manage friction
Friction complicates the clean torque balance assumed by the model. Contact friction acts in one direction while the robot pushes and reverses when it retracts, creating a separation between the two measured force curves. The method does not explicitly identify every frictional parameter. Instead, it combines the opposing push and retract measurements so that their biases partly counteract each other.
That approach was especially clear with the hollow acrylic reference box, whose center of mass was designed to lie at its geometric center. At the least conservative tested safety setting, the authors report a mass estimate of 0.641 kilograms against a measured 0.658 kilograms. The estimated center-of-mass height was 14.594 centimeters against a reference value of 14.624 centimeters, and the predicted toppling angle was 17.566 degrees against 17.532 degrees.
Those differences correspond to reported relative errors of 2.618 percent for mass, 0.205 percent for center-of-mass height and 0.195 percent for toppling angle. The precision should be read in the context of a favorable reference object with flat surfaces and a stable pivot edge. It shows that the force-angle signal can contain useful inertial information under the experiment’s assumptions. It does not establish the same precision for irregular, flexible or moving objects.
Stopping early creates a measurement tradeoff
The central design problem is that the robot seeks information about the toppling point while trying not to reach it. The researchers address this with a configurable safety margin based on measured force. The robot stops its push after the resisting force falls to a selected fraction of the initial peak, using that decline as a proxy for approaching marginal stability.
A larger safety-margin value makes the interaction more conservative because the robot ends the push earlier. That reduces exposure to unstable motion, but it also removes the most informative part of the trajectory near the tipping threshold. The model must then extrapolate farther to infer the critical angle and center-of-mass height. In the reported tests, settings below 0.65 produced usable trajectories, while larger values truncated the measurements too early for reliable fitting.
Less conservative settings collect more of the force decline and can improve estimation when the object behaves as expected. They also bring the system closer to pivot changes, slipping and local deformation that violate its simplified mechanics. The authors found no single safety setting that was best for every object. Mass estimates were comparatively stable, while center-of-mass height was more sensitive to early stopping and measurement noise because it depends heavily on the inferred tipping angle.
Four objects show both range and failure
Beyond the acrylic box, the researchers tested a 3D-printed heart-shaped prism, a handheld flashlight and a computer monitor. Their measured masses ranged from 0.219 kilograms for the heart to 5.040 kilograms for the monitor. The authors report relative errors below approximately 5 percent for valid estimates across the tested objects and settings. Computation was quick once data had been collected, with reported fitting times measured in milliseconds. Physical interaction took longer, from roughly five seconds to more than 31 seconds depending on the object and safety margin.
The flashlight exposed a material limitation. At the least conservative setting, its round base shifted as the robot pushed, producing yaw and unstable motion rather than rotation about one fixed edge. That trial yielded no valid estimate. The failure matters because the model assumes a stable pivot, quasistatic motion and sufficiently consistent contact. Curved bottoms, compliant packaging, wheels, feet, cables or uneven surfaces could create similarly difficult behavior.
The test environment also supplied information that a field robot would need to obtain for itself. AprilTags made the object’s angle easy to track, the planar center-of-mass projection was treated as available, and researchers selected the push height. The authors identify removing the marker requirement and choosing informative pushes autonomously as future work. The study did not evaluate clutter, uncertain support surfaces or repeated probing of fragile objects.
A useful measurement primitive, not deployment proof
The practical value lies in giving robots another way to inspect objects before committing to a grasp or lift. A manipulator might eventually use a controlled nudge to choose a safer grasp location, predict how a bulky item will respond to a push or decide that an object should not be moved with the available hardware. This is particularly relevant when visual shape alone says little about an internal weight distribution.
That promise remains narrower than autonomous warehouse handling. The experiment used one industrial arm, one sensing arrangement and four objects under controlled conditions. It provides no independent replication, broad object benchmark or demonstration that the estimates improved a subsequent manipulation task. Its safety language concerns remaining below a modeled toppling threshold, not a general guarantee against damage or workplace hazards.
Even within those limits, the study advances a useful idea: physical interaction can be a form of perception. Instead of demanding a complete object model from vision, a robot can collect mechanical evidence through a deliberately constrained motion. The documented flashlight failure makes the result more informative, not less. It identifies the boundary between a promising force-guided measurement and the varied contact mechanics that practical systems must still handle.