AI Robotics ·
After a 94% tape-pick result, can factories rely on fewer force sensors?
On November 10 at the Conference on Robot Learning, University of Michigan researchers presented PRISM, which learns interactions among signals such as joint positions, velocities and commands. A physical robot arm achieved a 94% success rate in a 50-attempt tape-picking test, compared with 68% for a conventional neural network. [1][2][3]
Problem definition and research question
For a lightweight, consistent part, signals from a robot’s existing joints may be enough to improve its movement. In precision assembly, changes in clearance and friction can alter the load on a part even when its position looks right. Which factory measures should determine whether to spend engineering time on the control model or invest first in force sensing and validation equipment?
Evidence and causal analysis
PRISM changes the network’s input layer to represent interactions among physical variables. The tape-picking comparison used the same training demonstrations. Reported reductions in contact force and locomotion results came from simulations. Predicting damage or insertion success for different parts requires evidence across tolerances, gripper wear and extended use. In a different test, KAIST’s DiSPo guided a clamp through a 2.5-mm radial clearance using a setup that included cameras and a force/torque sensor. The tasks and setups differ, so the two studies’ success rates cannot isolate the contribution of force sensing. [2][3][4]
| Dimension | International evidence | Korean evidence and application |
|---|---|---|
| Task | Picking up a tape roll [2][3] | Passing a clamp through a tight clearance and pressing a button [4] |
| Model change | Representation of interactions in existing inputs [3] | Variable movement steps for different levels of precision [4] |
| Factory evidence to collect | Damage rates and contact-force distribution | Sensor upkeep and acceptance rates across part variation |
Comparison of alternatives
For low-load picking, automation engineering could retain the current gripper and safety controls while trying a replacement input layer. That limits new hardware and makes comparisons easier, while requiring data collection and validation after retraining. Gripper wear could still become difficult to diagnose. For an insertion where contact forces determine quality, the equipment team could add a force/torque sensor and use explicit contact control. It could then review load changes and diagnose rework, at the cost of calibration, wiring maintenance and validation after replacement.
Implementation and evaluation
Compare the existing controller, the learned input change and a force-sensing controller with the same arm and gripper, matching part lots and time in service. Quality should log excessive contact, hidden damage, retries and 95th-percentile cycle time alongside task success. Independent instruments can measure physical load without feeding force signals to the learned policy. Any breach of the site-approved contact load stops the test. Pause expansion if the savings disappear when calibration and retraining are included, or if the new approach misses the acceptance threshold.
Conclusion and open questions
When the same robot picks parts and performs precision insertions, the cost of a single damaged part can change the comparison between sensor maintenance and model development. Which operations warrant their own force measurements?