Collect force-related data when the failure you need to explain is compression or slip that gripper position alone cannot distinguish. Then compare policies on the same objects, episodes and damage criteria. Adding a force field to a dataset is useful only if its measurement meaning and its contribution to the task remain visible.
The Just Add Force project, presented at a 2024 CoRL workshop, reports 130 successful grasp trajectories across 30 objects. Its setup estimates contact force from gripper current, records demonstrations at 5 Hz, and evaluates policies during the stationary grasp portion. The authors compare force-bearing and position-only diffusion policies and acknowledge noisy force estimates and remaining excessive force. Those results support investigation of that bounded grasp problem, not a claim of safe handling across arbitrary objects or transport motions.
Begin with the failure that width cannot explain
Imagine two grasps reaching the same final opening. One holds a package securely; the other deforms it. Width records the geometry commanded or observed at the fingers, but the object response can differ. A contact-related signal may help separate those outcomes.
Before collecting a new dataset, list the failure labels the intended application needs: unacceptable deformation, slip, dropped object, incomplete closure, or uncertain contact. Define them for the application rather than importing a research video's visual judgment. A favorable grasp score is not meaningful if damage is excluded from its definition.
Do not collect only successful examples if the planned evaluator needs to diagnose failures. Keep the labels and the reason for each disposition. A dataset rich in successful closures but poor in troublesome contacts can hide the situations that motivated force sensing.
A current-derived force estimate needs its own record
Motor current is a proxy, not a direct universal measurement of object contact force. The gripper's mechanics and operating state matter. Preserve the conversion method, actuator configuration, timestamp, position and available calibration information alongside the signal. If those inputs change between recording and deployment, the numerical meaning may change too.
The research setup's recording rate is a property of that experiment. It is not a recommendation for every gripper or contact event. For the proposed application, determine whether the recorded stream resolves the changes needed by the learner and evaluator. Do not infer a control-loop rate from a dataset sampling rate.
The site's visual and tactile sensor-fusion guide provides adjacent context for joining observations. Here the selection question is narrower: whether a traceable contact-related variable improves the grasp decision.
Budget for labels and validation as well as the signal
A fair application comparison uses the same training episodes, held-out objects, input preprocessing and task definition for force-bearing and position-only variants. Report damage and slip separately; their tradeoff can be hidden by one combined success percentage.
- Reserve held-out objects or recording sessions that represent the intended variation.
- State how deformation is judged and who owns that judgment.
- Record whether the object remains held during the relevant post-grasp motion.
- Compare the result with the force input removed or disturbed, so the evaluator can see whether the policy actually uses it.
- Include failures and uncertainty, rather than treating missing observations as successful contact.
These are proposed evaluation choices, not experiments performed here. They also expose the costs: labeled failure episodes, signal characterization, synchronization, model training and application acceptance. A method that uses no additional force sensor still incurs engineering work.
Choose force-data collection when that work addresses a defined gap in the position-only baseline. If the dominant failures come from localization or an unsuitable contact surface, collecting more force traces may leave the real problem unchanged. If the policy improves only stationary grasping, retain that boundary until additional motion has been evaluated.
Source checked October 14, 2026. No model was trained and no physical grasp or damage test was performed for this article.