During manipulation, the dexterous hand makes contact with the object, generating contact forces that, in turn, induce changes in joint torques. This paper estimates the contact forces from joint torque measurements, thereby eliminating the need for tactile sensors.

Conventional IL constructs a policy through a neural network, mapping from states and contact force  to actions  . We propose two novel IL frameworks, namely DexSensorless-Ident and DexSensorless-CFA. Ident represents identification-based contact force estimation, and CFA refers to the proposed Contact Force Autoregressive module. DexSensorless-Ident employs estimated contact forces from dynamic model as additional inputs. DexSensorless-CFA trains the CFA module using dynamically estimated contact forces,  internalizing the robot’s dynamics as model parameters to bridge the gap between analytical modeling and data-driven learning.