Torque-change alignment transfers a simulated grasp policy to a direct-drive hand
Sogang University researchers calibrate motor current, subtract consecutive torque readings and inject measured noise. The hand succeeds in 208 of 210 object trials without vision or tactile sensors.

Sogang University researchers released a torque-observation alignment method on September 24, 2026 for zero-shot transfer to a multifingered direct-drive gripper. Instead of adding tactile sensors, the method turns motor current into a contact cue: it calibrates the torque scale, differences consecutive readings to suppress offset, and injects measured noise during simulation training. The deployed policy then uses only joint positions and torque changes. Paper and version record.
Explore tactile sensing for the broader comparison.
Key takeaways
- The complete method records 100% success on nine in-distribution object variants, with ten physical trials per object.
- Across 21 objects and 210 trials, it succeeds 208 times: 100% on nine training-shape objects and 98.3% on 12 objects with unseen geometry, material or surface properties.
- Torque differencing removes a slowly varying baseline, but also hides sustained absolute load. The result should not be read as a replacement for calibrated force or tactile sensing in every task.
What changed
Direct-drive actuators expose interaction through motor current, but a policy trained on ideal simulated torque sees the wrong distribution on hardware. The paper separates that gap into scale, offset and noise.
First, a dynamometer estimates a motor-type-specific current-to-torque constant. Second, both simulator and robot feed the policy the change in torque from one 50 ms step to the next rather than the absolute reading. If the hardware bias changes slowly, subtraction cancels most of it. Third, Gaussian noise fitted from the dynamometer measurements is injected during training.
A privileged teacher learns in simulation, then a student policy is distilled for deployment. Training uses three object families—cuboids, cylinders and spheres—with three sizes each. The student runs at 20 Hz and controls a three-finger, nine-degree-of-freedom direct-drive gripper through grasp, lift and hold stages.
Results under the reported conditions
The ablation evaluates five observation variants on nine in-distribution objects, ten trials per object and condition. A position-only student reaches 15.6% on hardware despite 68.9% simulated success. Absolute torque with noise reaches 26.7%. Scale calibration plus differencing without injected noise reaches 62.2%; only the full scale, difference and noise condition reaches 100%. These are observation-alignment variants, some sharing trained policies, not five independent training methods. Ablation and protocol.
The broader evaluation adds 12 out-of-distribution household objects. The full method succeeds on all 90 in-distribution trials and 118 of 120 out-of-distribution trials, for 208 of 210 overall. The two failures are a light bulb and a tennis ball; the light bulb tilts and falls during holding, and the tennis ball slips during lifting. The separate ablation discussion of wrong-scale overtravel and noiseless oscillation should not be substituted for these two full-method failures.
These are repeated trials from a prescribed non-contact starting condition on a fixed gripper. The objects are within its workspace, and the test does not include arm motion, clutter, deformable-object safety or arbitrary initial hand-object geometry.
What this means for robotics
RoboSkin analysis: for direct-drive hands, actuator telemetry can serve as a low-cost contact channel if the learning pipeline respects how that telemetry changes between simulation and hardware. Scale calibration and temporal differencing are simple enough to inspect, unlike a learned black-box adaptation layer.
The result also draws a useful boundary between proprioceptive contact detection and robot skin. Torque changes can reveal contact onset and changing load throughout the mechanism, but they do not localize pressure across a surface or preserve static force. Engineers should choose between them based on the task state that must be observed.
Limitations and availability
This is an arXiv v1 preprint, not an independently reproduced system. Training runs for 2,000 iterations on an RTX 5090, but the paper does not establish that GPU as a minimum requirement. Calibration is tied to the tested motor type and dynamometer conditions; variation across individual motors and operating temperatures remains unknown.
The authors explicitly note that differenced torque suppresses sustained load and assumes slowly varying offsets. Faster dynamics may need more compensation. No official public code, trained policy, calibration dataset, CAD package or software license was verified in this review. The Sogang RIM Lab page lists the publication, but listing a paper is not an implementation release.
Which alignment operation addresses which error?
A current-to-torque fit does not supply a surface contact map. Keep the three operations separate when assessing whether the method fits your hand:
| Operation | What the paper changes | What remains to check on another hand |
|---|---|---|
| Scale calibration | Fits motor current to torque using a dynamometer | Motor instance, temperature, load range and the reference used |
| Consecutive differencing | Supplies torque change rather than absolute torque | Lost static-load information and whether offsets vary slowly enough |
| Noise injection | Adds measurement-informed variation during training | Noise under your acquisition, filtering and operating conditions |
The last column is an integration checklist, not a demonstrated transfer result. Paper methods and limitations support the distinction between calibration and policy training: zero-shot policy transfer still uses physical calibration evidence. Compare this indirect signal with measured tactile contact and controller inputs.
Sources and related resources
- arXiv v1 record, submitted September 24, 2026
- Full method, 210-trial evaluation and limitations
- Sogang RIM Lab publication list
- RoboSkin guide to full-hand sensing
*Evidence review updated November 4, 2026. The original paper and article dates are retained; source inspection is not a RoboSkin.ai reproduction.*


