Improving Current-based Torque Sensing forHarmonic Drives Leveraging Asymmetry inDirectional Efficiency.

Torque estimation is a critical requirement for several manipulation and mobility tasks such as closed-loop force control, gravity compensation, contact detection, etc. This can be achieved by direct sensing with torque sensors which are heavy and significantly increase the load requirements of the system. An alternative approach is to using indirect sensing which estimates torques by measuring current consumption of the actuator. Unfortunately, this indirect approach suffers from low accuracy. We propose an estimation algorithm to improve this accuracy for the popular harmonic drive-based actuators. This is achieved by exploiting the asymmetry in efficiency which depends on the direction of torque transmission i.e. the efficiency of torque transmission is significantly lower in the output to input direction compared to input to output. This direction can be identified from the motor current and encoder data when the actuator has non-zero angular velocity but is not observable when the actuator is not moving. Our proposed solution significantly improves the accuracy for the former case and provides upper and lower bounds of torque estimates for the latter case. This is validated through experiments on hardware and comparison with ground truth from direct torque sensing.