Bidirectional recurrent learning of inverse dynamic models for robots with elastic joints: a real-time real-world implementation

Collaborative robots, or cobots, are designed to work alongside humans and to
alleviate their physical burdens, such as lifting heavy objects or performing tedious
tasks. Ensuring the safety of human–robot interaction (HRI) is paramount for
effective collaboration. To achieve this, it is essential to have a reliable dynamic
model of the cobot that enables the implementation of torque control strategies.
These strategies aim to achieve accurate motion while minimizing the amount of
torque exerted by the robot. However, modeling the complex non-linear dynamics
of cobots with elastic actuators poses a challenge for traditional analytical
modeling techniques. Instead, cobot dynamic modeling needs to be learned
through data-driven approaches, rather than analytical equation-driven modeling.
In this study, we propose and evaluate three machine learning (ML) approaches
based on bidirectional recurrent neural networks (BRNNs) for learning the inverse
dynamic model of a cobot equipped with elastic actuators. We also provide
our ML approaches with a representative training dataset of the cobot’s joint
positions, velocities, and corresponding torque values. The first ML approach uses
a non-parametric configuration, while the other two implement semi-parametric
configurations. All three ML approaches outperform the rigid-bodied dynamic
model provided by the cobot’s manufacturer in terms of torque precision while
maintaining their generalization capabilities and real-time operation due to the
optimized sample dataset size and network dimensions. Despite the similarity in
torque estimation of these three configurations, the non-parametric configuration
was specifically designed for worst-case scenarios where the robot dynamics are
completely unknown. Finally, we validate the applicability of our ML approaches by
integrating the worst-case non-parametric configuration as a controller within a
feedforward loop. We verify the accuracy of the learned inverse dynamic model by
comparing it to the actual cobot performance. Our non-parametric architecture
outperforms the robot’s default factory position controller in terms of accuracy.