A Differentiable Newton–Euler Algorithm for Real-World Robotics

The identification of dynamical systems from data is a powerful tool in robotics [3]. Learned analytic models may be used for control synthesis and can utilized gravitational inertial compensation [75]. When as simulators, these reduce the sample complexity data-driven methods such RL [15, 11, 28, 46]. For applications, where out-of-distribution prediction typically required, ability to generalize beyond acquired critical. Any modeling error exploited by methods, exploitation result failure on physical system. To ensure sufficient out-of-sample generalization, model’s hypothesis space an important consideration. Ideally, this should defined that only plausible trajectories, are physically consistent have bounded energy, generated.