Calibrating the torque constant with a dynamometer and aligning torque-difference observations lets a direct-drive multifingered gripper reach 100% zero-shot sim-to-real grasp success on nine in-distribution objects

It introduces a torque-observation alignment pipeline for direct-drive actuators that treats the scale, offset, and noise gaps between simulated and real torque separately. Relative to the common practice of using raw torque tau(t) as the observation, the method assigns the scale gap to the dynamometer-calibrated K_tau*, the domain-dependent constant bias to the torque difference delta_tau(t) = tau(t) - tau(t-1), and the noise gap to additive Gaussian noise taken from dynamometer measurement data. The pipeline is presented through method description and an ablation study comparing it with alternative alignment variants on nine in-distribution objects; the loaded text is summary-level and does not report calibration curves, noise parameters, or per-object results.

A teacher-student grasping policy is trained entirely in simulation and the distilled student is deployed zero-shot on a multifingered DD gripper, relying only on joint positions and torque differences for proprioceptive grasping. This validation embeds the alignment method in a complete sim-to-real transfer chain rather than stopping at the observation level, indicating that the aligned torque-difference observation is sufficient to drive grasping decisions on real hardware. Evidence comes from deployment on a real multifingered DD gripper with policy inputs restricted to joint positions and torque differences; the text does not report training scale, distillation details, or the number of real trials.

In the ablation study on nine in-distribution objects, the proposed alignment method achieves 100% grasp success. The result links the choice of alignment scheme directly to grasp success, suggesting that handling scale, offset, and noise together underpins the robustness of zero-shot transfer. Evidence is the grasp success rate from the ablation on nine in-distribution objects; the text does not give the specific numbers for the alternative variants, confidence intervals, or out-of-distribution results.

The result targets robots with direct-drive actuators whose motor current maps linearly to joint torque through a motor-type-specific torque constant K_tau, and it presupposes dynamometer calibration and measured noise statistics. It enables follow-up work to train in simulation and deploy zero-shot on real DD hardware a proprioceptive grasping policy that uses only joint positions and torque differences, and it offers a reproducible alignment path for bringing torque-based observations into sim-to-real transfer. The applicable setting is the multifingered DD gripper and in-distribution grasping task validated here.

A careful reader would still want to know the individual success rates of the alternative alignment variants and how many trials, under what success criterion, produced the 100% figure; how the method behaves on out-of-distribution objects or new tasks beyond the nine in-distribution objects; whether the calibrated K_tau* and noise statistics need recalibration across motors, payloads, or long operation; and whether the approach still applies to non-direct-drive transmissions or joints with pronounced friction and compliance. Because the loaded text is summary-level and lacks figures and itemized experimental data, these questions cannot be answered from the available material.