PaperlayerView exact revision on arXiv

Source-linked AI summary

Simple Torque-Observation Alignment for Zero-Shot Sim-to-Real Grasping with a Direct-Drive Gripper

Doyoung Kim, Edgar Lee, Hyeonsun Park, Chunghyeon Lee, Chihyun Han, Uisu Hwang, Seokhwan Jeong

arXiv:2609.29031v1cs.RO

TL;DR

Simulated and measured torque observations can differ in scale, offset, and noise, complicating torque-based sim-to-real grasping. The paper aligns observations and trains a teacher–student policy in simulation; the deployed method achieves 100.0% success on nine ID objects and 99.0% overall across 21 objects. Its evaluated scope is single-object grasping with a fixed-base gripper under prescribed initial conditions.

  • Problem

    Simulated and real torque observations can differ in scale, additive bias, and noise, making reliable sim-to-real policy transfer challenging.

    Source: §I · p. 1

  • Method

    The method calibrates torque scale, uses torque differences to suppress offsets, and injects measurement-informed Gaussian noise into simulation training before teacher–student deployment.

    Source: §I · p. 2, B. Simple Torque Observation Alignment · p. 3, §V · p. 8

  • Results

    The method achieves 100.0% success on nine ID objects and 98.3% on twelve OOD objects, with 99.0% overall success.

    Source: §V · p. 8

  • Takeaways & Limitations

    Under evaluated grasp–lift–hold conditions, joint positions and torque differences support zero-shot proprioceptive grasping without vision or dedicated tactile or force/torque sensors.

    Source: §V · p. 8, §IV · p. 8

    Experiments cover single-object grasping with a fixed-base gripper under prescribed initial conditions and within-workspace objects.

    Source: §IV · p. 8

Abstract

from arXiv · show

Torque observations in reinforcement learning remain challenging because simulated and measured torque differ in scale, offset, and noise. In this paper, we propose a simple torque observation alignment method for robots with direct-drive (DD) actuators, in which motor current maps linearly to joint torque through a motor-type-specific torque constant K_tau. First, dynamometer calibration identifies K_tau* and corrects the scale mismatch between simulated and real torque. Second, the method uses delta_tau(t) = tau(t) - tau(t-1) as the observation in both domains to eliminate the constant offset instead of using the direct torque tau(t), which carries a domain-dependent bias. Third, Gaussian noise obtained from the dynamometer measurement data is injected during the learning process. To validate the proposed method, we train a teacher-student grasping policy entirely in simulation and deploy the distilled student on a multifingered DD gripper. The deployed policy performs proprioceptive grasping using only joint positions and torque differences. We conduct an ablation study comparing the proposed method with alternative alignment variants on nine in-distribution (ID) objects. The proposed method achieves 100% grasp success. These results demonstrate that the proposed alignment method improves the robustness of zero-shot policy transfer on the DD gripper against real-world torque-observation mismatches.

I. INTRODUCTION

The introduction motivates torque feedback for proprioceptive grasping while noting that simulated and measured torque differ in scale, baseline, and noise. It presents a calibrated, differenced torque-observation pipeline and simulation-trained student policy, reporting high real-world grasp success.

I. INTRODUCTION

  • Simulated and real torque observations differ in scale, additive bias, and noise, potentially changing how a simulation-trained policy interprets contact and load changes.Torque feedback can complement joint positions with information about contact onset and changing grasp load.Source: §I · p. 1
  • Prior approaches use randomization or actuator models, while related observation alignment uses fingertip force measurements and tactile feedback; this method avoids tactile sensors at deployment.The cited alternatives either do not explicitly calibrate correspondence between domains or require actuator-specific identification or model training.Source: §I · pp. 1–2
  • The method calibrates motor-type-specific torque constants, uses torque differences to suppress offsets, and injects measurement-informed Gaussian noise during training.Datasheet-based estimates exceed dynamometer-calibrated values by approximately 38.0% and 7.8% for the two motor types.Source: §I · p. 2
  • The framework combines torque-observation alignment with simulation-based teacher–student learning for zero-shot deployment using joint positions and torque differences.The policy is trained entirely in simulation and deployed without vision or dedicated contact sensors.Source: §I · p. 2
  • In real-world zero-shot grasp–lift–hold experiments under evaluated conditions, the proposed method achieved 100.0% success on nine in-distribution objects and 99.0% overall across 21 objects, including twelve previously unseen out-of-distribution objects.Source: §I · p. 2

II. METHOD

The method has two stages: align torque observations across simulation and reality, then train a simulation teacher and distill it into a deployment student.

II. METHOD

  • The method first aligns joint-torque observations across simulation and real-world domains.This is the first of the two pipeline stages.Source: §II · p. 2
  • A simulation-trained teacher uses privileged information and aligned torque observations for the grasping task.The teacher policy is trained in simulation.Source: §II · p. 2
  • The teacher is distilled into a student policy for real-world deployment.Distillation forms the pipeline’s second stage.Source: §II · p. 2

A. Direct-Drive Gripper Platform

The platform is a three-fingered direct-drive gripper, with motor-current measurements converted to joint-torque estimates; the figures show its layout and dynamometer calibration setup.

A. Direct-Drive Gripper Platform

  • The platform is a three-fingered, 9-DoF gripper with direct-drive actuators integrated into its fingers.Figure 3 presents the gripper, actuator integration in one finger, and joint-and-link configuration.Source: A. Direct-Drive Gripper Platform · p. 2, Figure 3 · p. 3
  • Figure 4 shows dynamometer calibration using a test motor, load motor, and inline torque sensor, with torque-versus-current plots for GL40 and GL60.Dashed lines denote linear-regression fits.Source: Figure 4 · p. 3
  • Each finger has three independently actuated joints, and measured quadrature current is converted to joint-torque estimates using the corresponding torque constant.The MCP joint uses a GL60 motor; the PIP and DIP joints use GL40 motors.Source: A. Direct-Drive Gripper Platform · p. 3

B. Simple Torque Observation Alignment

The alignment method models real torque observations as scaled simulated torque plus baseline and residual error. It calibrates the scale, differences observations over time to reduce baseline mismatch, and injects measurement-derived noise during training.

B. Simple Torque Observation Alignment

  • The cross-domain model represents real current-derived torque as scaled simulated torque with a slowly varying baseline and residual noise.The baseline includes gravity, friction, and current offset; residual noise reflects measurements and unmodeled actuator effects.Source: B. Simple Torque Observation Alignment · p. 3
  • Dynamometer calibration identifies effective torque constants for current-to-torque conversion, reducing scale mismatch between simulated and real observations.The identified constants and regression accuracy are summarized in Table I.Source: B. Simple Torque Observation Alignment · p. 3, Table I · p. 3
  • Temporal differencing cancels a constant baseline mismatch and attenuates slowly varying components while preserving cues about interaction-load changes.The same differencing process is applied to observations in both domains.Source: B. Simple Torque Observation Alignment · p. 4
  • During teacher training, Gaussian noise derived from dynamometer regression RMSE is injected into simulated torque observations using motor-type-specific standard deviations.The assigned standard deviations are 0.009843 N m for MCP joints and 0.002305 N m for PIP and DIP joints.Source: B. Simple Torque Observation Alignment · p. 4

C. Learning Method

The method trains a simulation teacher with privileged information and distills its behavior into a student that deploys using joint positions and torque differences. Training combines task rewards and action penalties with randomized initial conditions, physical parameters, and disturbances.

C. Learning Method

  • Training objects are cuboids, cylinders, and spheres, each with three size variants, while domain randomization varies initial conditions, physical parameters, and external disturbances.Object poses and joint-angle offsets are resampled each episode; mass, actuator stiffness and damping, and friction are randomized at initialization.Source: C. Learning Method · pp. 4–5
  • The teacher is trained with PPO in 9,000 parallel simulation environments for 2,000 iterations, and the student is trained by behavior cloning for 2,000 epochs.For each student configuration, 90,000 demonstration trajectories are collected; teacher rollouts are retained according to final-step reward thresholds.Source: C. Learning Method · p. 4
  • The deployable student uses only joint positions and torque differences, while the teacher also receives simulator-only privileged information.The teacher is trained with PPO; the student is trained by behavior cloning on teacher demonstrations.Source: C. Learning Method · pp. 4–5
  • The reward combines a lifting-and-centering task objective with penalties for action magnitude and abrupt changes between consecutive commands.The action is a 9-dimensional vector of relative joint-position changes, and control runs at 20 Hz.Source: C. Learning Method · p. 5

III. EXPERIMENTS

The experiments assess alignment variants, grasping across in-distribution and unseen objects, and representative failure patterns. All policies are trained in simulation and transferred without fine-tuning or adaptation.

III. EXPERIMENTS

  • The evaluation compares five alignment variants on nine in-distribution objects, tests the proposed method on 21 objects, and examines grasping sequences and torque-change trajectories.The 21-object set includes nine ID objects and twelve OOD objects with unseen shapes, materials, and surface properties.Source: §III · p. 5, Figure 7 · p. 6
  • Table VI summarizes simulation and real-world success rates for the student policies.The ablation includes a position-only policy to assess torque feedback beyond joint positions.Source: §III · p. 5, Table VI · p. 6
  • All policies are trained entirely in simulation and transferred to the real world without fine-tuning or adaptation.The evaluation uses only actuator proprioception on a direct-drive multifingered gripper.Source: §III · pp. 5–6

A. Ablation Study

Across the nine in-distribution objects, the complete alignment method achieves 100.0% success, while ablations show lower success for variants missing alignment operations. The comparisons examine calibration, temporal differencing, and measurement-derived noise injection.

A. Ablation Study

  • The five variants combine datasheet or calibrated torque constants, absolute torque or torque differences, and measurement-derived noise injection in different ways.Noise Only and Scale + Noise use absolute torque; Bias + Noise and Our Method use torque differences with noise, while Scale + Bias uses torque differences without noise.Source: A. Ablation Study · p. 6
  • 100.0%: the proposed method succeeds in the real world, compared with 26.7% for the absolute-torque policy and 62.2% for the torque-difference policy without noise injection.All three torque-based students achieve 100.0% success in simulation; the position-only student achieves 68.9% in simulation and 15.6% in the real world.Source: A. Ablation Study · p. 6
  • 100.0%: Our Method succeeds on all nine ID objects in the real-world ablation.Fig. 6 reports observed success rates for the three alignment operations across nine objects, with ten trials per object and variant.Source: A. Ablation Study · p. 6
  • Adding scale calibration raises success from 55.6% to 100.0%, while adding noise injection to Scale + Bias raises success from 62.2% to 100.0%.The comparisons support complementary roles for temporal differencing, scale calibration, and noise injection.Source: A. Ablation Study · p. 7

B. Real-World Grasping Experiments

The proposed method achieves 100.0% average success on ID objects, 98.3% on OOD objects, and 99.0% overall. The two unsuccessful trials involved a light bulb during holding and a tennis ball during lifting.

B. Real-World Grasping Experiments

  • The evaluation covers twelve previously unseen OOD objects with combinations of geometry, material, and surface properties not represented during training.Each object is evaluated over ten trials from the same prescribed initial condition.Source: B. Real-World Grasping Experiments · p. 7
  • The method achieved average success rates of 100.0% on ID objects and 98.3% on OOD objects, resulting in an overall success rate of 99.0%.The two unsuccessful trials involved a light bulb that tilted and fell during holding and a tennis ball that slipped between the fingers during lifting.Source: B. Real-World Grasping Experiments · p. 7
  • The deployed policy uses joint positions and motor-current-derived torque changes, without external sensors, for zero-shot grasping of unseen objects.Direct-drive actuation allows torque changes to provide cues about contact onset and changing interaction loads.Source: B. Real-World Grasping Experiments · p. 7

C. Failure-Mode Analysis

The failure-mode analysis links mismatched torque observations to unbalanced three-finger support, while the complete method stabilizes grasping through grasp establishment, lifting, and holding.

C. Failure-Mode Analysis

  • Datasheet torque constants overestimate GL60 and GL40 channels by 38.0% and 7.8%, respectively, altering relative joint torques and accompanying middle-finger overtravel.The calibrated variant without training noise instead shows repeated action corrections and unstable torque redistribution before lifting.Source: C. Failure-Mode Analysis · p. 7
  • After the initial closure transient, the complete method stabilizes coordinated torque transitions, whereas both ablated variants continue redistributing torque.The analyzed variants are Our Method, Scale + Bias, and Bias + Noise.Source: C. Failure-Mode Analysis · p. 7
  • Both observation mismatches are associated with loss of balanced three-finger support, while Our Method maintains coordinated support through grasping, lifting, and holding.Grasp-establishment failures lack stable three-finger contact; lift failures cannot support the object, and hold failures lose support after lifting.Source: C. Failure-Mode Analysis · p. 7

IV. DISCUSSION AND LIMITATIONS

The discussion reports proprioceptive grasping of unseen objects without vision or dedicated tactile or force/torque sensors, while identifying limits in torque feedback, parameter transfer, and evaluation scope.

IV. DISCUSSION AND LIMITATIONS

  • Torque differences do not directly observe sustained torque levels, and suppressing slowly varying offsets may require added compensation in more dynamic tasks.The evaluated lift is dynamic, whereas settled holding is quasi-static.Source: §IV · p. 8
  • Torque constants and noise parameters were identified under specific dynamometer conditions; transfer across motor units and operating temperatures remains unevaluated.The experiments also use single-object grasping with a fixed-base gripper, prescribed initial conditions, and objects within its workspace; arm-mounted evaluation is future work.Source: §IV · p. 8
  • The policy grasps unseen objects using joint positions and torque changes without vision or dedicated tactile and force/torque sensors.The paper presents proprioceptive feedback as a useful source of contact information for direct-drive grippers.Source: §IV · p. 8

V. CONCLUSION

The conclusion reports high zero-shot grasp success using a three-part torque-alignment method that needs neither a learned actuator model nor real-world policy training. Under the evaluated grasp–lift–hold conditions, joint positions and torque differences support proprioceptive grasping without vision or dedicated contact sensors.

V. CONCLUSION

  • Dynamometer calibration reduces torque-scale mismatch, temporal differencing cancels constant offsets and attenuates slowly varying baselines, and dynamometer-residual Gaussian noise is injected during training.The method requires neither a learned actuator model nor real-world policy training.Source: §V · p. 8
  • 100.0% success on nine ID objects, 98.3% on twelve OOD objects, and 99.0% overall were achieved when all three alignment components were combined.The evaluated grasp–lift–hold policy uses joint positions and torque differences without vision or dedicated tactile or force/torque sensors.Source: §V · p. 8
  • The results support motor-current-derived torque changes as proprioceptive feedback about contact onset and varying interaction loads on direct-drive grippers.The paper concludes that joint positions and torque differences support zero-shot proprioceptive grasping under the evaluated conditions.Source: §V · p. 8
Loading 2609.29031v1…