Force-Based Object Dimension Estimation in Practical Bin Picking Scenarios
Tristan Fogt, Louis Hardinghaus, Georg Siegemund, Timon Adler, Holger Kunz, Arne Glodde, Sina Rahlfs, Franz Dietrich
Technical University of Berlin and FORMHAND Automation GmbH
Annals of Scientific Society for Assembly, Handling and Industrial Robotics 2024 (MHI 2024) — Springer · 2024

Introduction
Bin-picking systems commonly rely on cameras to determine object geometry and pose. Camera-based systems can, however, introduce additional hardware, calibration, computational, and integration requirements.
This work investigates whether a robot can estimate the dimensions of a grasped object through physical interaction alone.
At a glance
A camera-less bin-picking system may successfully detect and grasp objects using force feedback but still lack information about the geometry of the object it is handling.
After grasping an object, the robot probes it against known surfaces and determines its bounding dimensions from robot position and force/torque feedback.
The method predicted object dimensions within approximately 8–18%, depending on object geometry and measurement increment.
Measuring without an additional camera
The method builds on the robot and force/torque sensing already present in a force-feedback bin-picking process, rather than adding a dedicated vision system.
The grasped object itself is used as a probe against known references in the environment.
Determining the bounding dimensions
After an object is removed from the bin, the robot lowers it toward the work surface to determine one dimension, then moves it toward the bin edge to determine the remaining dimensions.
The robot then rotates the object and repeats the measurement; the resulting readings are combined into an estimate of the object's bounding box.
Accuracy versus process time
The study evaluated 90°, 45°, 30°, and 15° rotation increments. Smaller increments can provide a tighter estimate but require more probing operations and therefore increase process time.
Measured process time ranged from roughly 18 seconds for the 90° strategy to approximately 55 seconds for the 15° strategy.
Evaluation with different geometries
The approach was evaluated on three practical object groups — cardboard boxes, cans, and tennis balls — testing it across clearly different object shapes.
Practical relevance
The approach is less precise than dedicated vision-based dimensional measurement, but it obtains additional geometric information without adding another sensing system.
This is particularly relevant for camera-less or sensor-minimal bin-picking concepts, where every additional sensor adds cost, calibration effort, and integration complexity.
Highlights
- Object dimensions were estimated within roughly 8–18% deviation, depending on object geometry and rotation increment.
- Measurement time rose from about 18 s (90° increments) to about 55 s (15° increments).
- Going from 90° to 30° brought only a modest improvement; the clear gain came at 15°.
- The method uses the robot and force/torque sensor already present in a camera-less bin-picking cell.
Methodology
- Hardware: a UR10e robot with integrated force/torque sensor, commanded from Python over RTDE, and a 30 cm × 40 cm × 20 cm plastic bin.
- Height: the grasped object is lowered onto the work surface until a force threshold is exceeded; the gripper height minus the pre-recorded work-surface height gives the first dimension.
- Width and depth: the object is moved against the outside wall of the bin at 90° intervals and the gripper distance to the pre-recorded bin edge is measured; opposite measurements are added to obtain the axis-aligned bounding box.
- Finer estimates: the object is probed at 45°, 30°, and 15° increments to fit the second and third dimensions more tightly, at the cost of process time.
- Test objects: cardboard boxes, 330 ml cans, and tennis balls placed freely in the bin; each increment was tested with twenty objects per category.
- Evaluation: the volume calculated by the algorithm was compared with the manually measured volume of each object.
Findings
- From 90° to 15° the mean deviation fell from 16.19% to 10.16% for boxes, from 16.25% to 8.53% for cans, and from 13.19% to 10.31% for tennis balls.
- For boxes and cans the improvement from 30° to 15° was significant; for tennis balls it was not.
- For the tennis balls, the height dimension dominated the error (about 26–30%), while width and depth deviated by roughly 2–7%. The balls sank about 2 cm into the vacuum gripper.
- For cylinders, the spread of the deviation also shrank with finer increments.
- The tactile approach is less precise than camera-depth or RGB-D methods, but needs no extra equipment or calibration and far less computation.
Figures


Method diagram

Limitations
- The first dimension remains axis-aligned, so the result may not match the true minimum bounding dimensions of an object resting at an angle.
- Objects that deform or sink into the gripper, such as the tennis balls, distort the height measurement.
- Finer increments cost time: about 55 s per object at 15°, which suits dimensioning one object at the start of a task better than every object.
- The method was evaluated on three object groups in a single hardware setup.
Citation
Fogt, T., Hardinghaus, L., Siegemund, G., Adler, T., Kunz, H., Glodde, A., Rahlfs, S., & Dietrich, F. (2024). Force-Based Object Dimension Estimation in Practical Bin Picking Scenarios. In Annals of Scientific Society for Assembly, Handling and Industrial Robotics 2024 (MHI 2024). Springer. Published online 2026.
Interested in a research collaboration?
Get in touch to discuss joint research, data collection, or applied evaluation of new methods.