Non-parametric model (NID) feedforward control deplyment on the real robot

Video.1.MP4

The video depicts the Baxter controlled by the NID neural network in a feedforward loop. We compared the NID performance (accuracy) to a conventional PD. We also compared their performances under perturbations. We finally showed NID ability to track the reference trajectory in the absence of active feedback control.

Dataset Baxter robot

This dataset contains the training and test data that were used for the recurrent neural network to learn the inverse dynamics of the Baxter robot.

The data correspond to the angles and velocities of the joints when a torque is applied in order to follow a trajectory.

Trajectories

• Random: The desired trajectory is random points in the robot's workspace.

• sin_XYZ_(Tt)s_T(Txy)Tz(Tz)T_r(r): The desired trajectory is defined by sinusoids in x, y and z.

       * Tt: total period [seconds]  
       
       * Txy: Period in xy [seconds]  
       
       * Tz: period in z [Txy]  
       
       * r: radius in XY [cm]  
  

• spiral: The desired trajectory is a spiral with radius in XY depending on time [r(t) = kt].

Data description

1:7 joint angles, from s0 to w2

8:14 joint velocities, from s0 to w2

15:21 joint torques, from s0 to w2