An arXiv version-one preprint describes a robot-gripper controller that uses jaw-encoder position and motor effort rather than tactile or force-torque sensing. In MuJoCo simulation, the proposed controller recorded at least 98% grasp success with 0% damage across medium-to-firm stiffnesses and the certified range of its user-set compression limit, ε. In hardware tests, it recorded 40% damage on soft 3D-printed TPU cubes at ε=0.7%, compared with 100% for each of two fixed-force baselines.
The result varied with stiffness. In the very-soft simulated regime, at k=2,000 N/m, limits of ε≤1.4% stayed damage-free but peaked at 48% grasp success. At ε=1.5%, grasp success was 40% with 6% damage; at ε=2.0%, grasp success was 18% and damage was 76%. On soft hardware cubes, damage rose from 40% at ε=0.7% to 80% at 1.3% and 100% at 2.0%.
How the controller estimates a compression limit
The study asks whether a gripper can stop at a user-specified compression-strain limit, written ε, using only encoder position and motor effort. Put simply, ε is the compression ceiling the controller is asked to respect. The system does not use tactile or force-torque sensing.
Those signals are read at 50 Hz. Contact is detected from the amplitude and persistence of the motor response. The controller then estimates compression from an effort-based force proxy divided by the stiffness lower bound and object diameter, applying a proportional command that closes the jaws toward ε or holds there.
The claimed bound is conditional: the paper states that true compression remains at or below ε when ε is above the contact-detection floor, provided the assumed stiffness lower bound is valid. The authors list online stiffness estimation as a next step.
Firmer objects produced the clearest simulation result
The MuJoCo simulation grid contained 2,000 trials, with 50 seeded runs per cell, and pooled cube diameters of 35, 40 and 45 millimetres. It compared the proposed controller with two fixed-force baselines. The baselines shared the proposed controller’s contact detector and lift trajectory and differed only in the stopping predicate.
Across medium-to-firm stiffnesses and the certified ε range, the proposed controller recorded at least 98% grasp success and 0% damage. Neither fixed-force baseline attained that regime. The percentages were descriptive results from seeded trials; no confidence intervals, inferential tests or formal uncertainty analysis were reported.
The very-soft cell showed a sharp trade-off between the two measures. At k=2,000 N/m, limits of ε≤1.4% stayed damage-free but reached at most 48% grasp success; the ε=1.5% cell recorded 40% grasp success and 6% damage, while the ε=2.0% cell recorded 18% grasp success and 76% damage.
Soft cubes remained the harder test
Hardware validation used regular 40-millimetre 3D-printed TPU cubes with 5%, 10% and 15% infill, labelled soft, medium and stiff. Each method–cube combination used five seeded trials, for 75 hardware trials in total. Damage was defined differently on the two platforms: simulation flagged Hookean strain above 1.5% sustained for at least five control ticks, while hardware damage was a visual score of at least 1.
At the 0.7% limit, the proposed controller’s soft-cube damage rate was 40%, compared with 100% for each fixed-force baseline. Its reported grip force was about 62 load units, versus 103 for the fixed-low and 119 for the fixed-high baseline. Medium and stiff cubes generally recorded 100% grasp success and 0% damage; the exception was the stiff-cube cell at ε=0.7%, where grasp success was 80%.
Across the soft-cube hardware cells, damage rose monotonically from 40% at ε=0.7% to 80% at ε=1.3% and 100% at ε=2.0%. These percentages come from five seeded trials per method–cube cell, and the analysis reports no interval estimates or inferential tests.
Closing speed sets a lower floor
Contact detection limits how low the strain bound can be set. The reported detection floor increased linearly with closing speed, with a fitted relationship of 0.10% + 0.093% × v, where v is closing speed in millimetres per second. In the paper’s framing, speed becomes a throughput–gentleness knob: faster closing comes with a higher minimum usable limit.
What remains untested
The authors identify grasping very soft objects with rigid pads as a main limitation. They list validation on real fruit, replacing the fixed diameter with an effective contact dimension for irregular produce, and online stiffness estimation as next steps.
The evidence is limited to controlled MuJoCo simulation and controlled SO-ARM101 experiments using regular TPU cubes. It does not establish performance on biological fruit or irregular produce. The paper also does not report confidence intervals, inferential tests or formal uncertainty analysis.
Because the simulation and hardware damage endpoints differ, their percentages are not directly interchangeable; neither endpoint establishes how biological fruit would fare.
The authors report that the simulation environment, controller implementation, hardware driver and experiment data are available through a linked GitHub repository. Whether the conditional bound and speed-dependent floor transfer beyond these controlled tests remains open.
Paper data and sources
Original title: Sensorless damage-safe grasping
Authors: Yusei Shuto, Danilo Vasconcellos Vargas
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text