The invention is notably directed to a computer-implemented method of controlling an autonomous
material handling machine (1). The
machine includes a lower structure (12), an upper structure (11), an arm (10) with arm joints actuated by arm joint actuators (461, 462), and an end
effector (110) suspended from the arm so that it can swing as a damped
pendulum. The arm (10) is pivotably connected to the upper structure, which can rotate relative to the lower structure through a slew joint (14), around a rotation axis of the slew joint, thanks to a slew motor (464). The method comprises repeatedly performing (S30) algorithmic cycles. Each algorithmic cycle comprises: updating (S31) state data (204), which comprise states of the arm joints and a
moment of inertia of the
material handling machine (1) with respect to the rotation axis, based on signals received (S20) from sensors of the
material handling machine (1); generating (S35), by a control policy (2065) trained with
reinforcement learning, control commands (203) for the slew motor (464) and the arm joint actuators (461, 462), based on the updated state data (204), and possibly a history (204h) thereof, a history (203h) of control commands previously generated by the control policy, and a target position constraint for the end
effector (110); and updating (S36) the history of control commands according to the control commands (203) generated last and instructing (S37) to form control signals based on the generated control commands to control (S40) the slew motor (464) and the arm joint actuators (461, 462). The invention is further directed to related systems and
computer program products.