Weight Sort
Lift a parcel and place it in the black container if light or the blue container if heavy.
Physical variation: Visually identical parcels have different masses.
Complete success: Correct destination, placement, release, and settling.
Benchmark tasks
Eight tasks. Three manipulation demands.
ForceBench evaluates how wrist force and torque support physical-property inference, resistance-dependent interaction, and contact-rich assembly under a shared simulation protocol.
2 property-inference tasks5 resistance-dependent tasks1 assembly task
Physical-property inference
The robot interacts with a parcel, infers its physical condition, and chooses the corresponding destination. Appearance is shared across conditions.
CoM Sort
Hold the parcel steadily, then place a centered parcel on the conveyor or an imbalanced parcel in the basket.
Physical variation: Identical appearance and total mass, with centered or offset centers of mass and different rotational loads about the grasp.
Complete success: Correct destination and placement, with release and settling.
Resistance-dependent interaction
The robot must adapt its motion direction, engagement, or stopping point to resistance encountered during interaction.
Faucet Turn
Open a valve against resistance that rises to a peak and then drops. Stop in the target region just beyond that transition.
Physical variation: The resistance-peak angle varies across instances.
Complete success: Reach the target region, release and withdraw, and leave the water on with the handle stationary.
Faucet Direction
Find the lower-resistance rotation direction and cross its resistance peak, leaving the valve open.
Physical variation: Both the lower-resistance direction and peak location vary.
Complete success: Cross the peak in the correct direction, release, withdraw, and leave the valve stationary. The final-angle criterion has a lower bound and no upper bound.
Stove Turn
Grasp and rotate a stove knob through an ignition resistance transition.
Physical variation: The transition occurs at different angles across instances.
Complete success: Cross the transition, reach the prescribed terminal angle, and activate the flame.
Stove Press
Depress a spring-loaded knob to permit rotation, then turn it through the ignition transition in the lower-resistance direction.
Physical variation: The direction and resistance-peak location vary.
Complete success: Coordinate axial engagement and rotation to achieve ignition.
Drawer Pull
Pull a drawer into a rising-resistance region and stop at the opening corresponding to a specified resistance level.
Physical variation: Resistance onset, extent, and breakaway resistance vary.
Complete success: Stop in the target opening range, release, and leave the drawer stable. Pulling beyond the mechanism’s limit is a failure.
Contact-rich assembly
Assembly tests constrained rotational insertion, complementing the property and resistance tasks.
Nut Thread
Align a grasped nut with a bolt, engage the threads, and rotate the nut down to a prescribed threaded height.
Physical variation: The target bolt position varies across instances.
Complete success: Reach the threaded target height while centered and upright, then release and withdraw. Under the task’s physical constraints, reaching this target assembly state implies correct thread engagement and screw-in insertion.
Physical decisions and complete execution
Success is evaluated using the full task-specific requirements. A correct destination, angle, or assembly state alone may not be enough: release, clearance, stability, and completion events also matter where specified.
The first two categories vary physical conditions without distinct visual appearances. These conditions can still affect visible motion and proprioception. The benchmark therefore measures the additional benefit of force, rather than assuming that force is the only possible source of information.
All tasks use a Franka arm with a parallel gripper, synchronized observations and actions, and seeded evaluation. See the protocol for the observation interface, controller, action chunks, and episode budget.