Sensor-assisted adaptive motor control under continuously varying context

Adaptive motor control under continuously varying context, like the inertia parameters of a manipulated object, is an active research area that lacks a satisfactory solution. Here, we present and compare three novel strategies for learning control under varying context and show how adding tactile sensors may ease this task. The first strategy uses only dynamics information to infer the unknown inertia parameters. It is based on a probabilistic generative model of the control torques, which are linear in the inertia parameters. We demonstrate this inference in the special case of a single continuous context variable - the mass of the manipulated object. In the second strategy, instead of torques, we use tactile forces to infer the mass in a similar way. Finally, the third strategy omits this inference - which may be infeasible if the latent space is multi-dimensional - and directly maps the state, state transitions, and tactile forces onto the control torques. The additional tactile input implicitly contains all control-torque relevant properties of the manipulated object. In simulation, we demonstrate that this direct mapping can provide accurate control torques under multiple varying context variables.