During delicate force-based contact manipulations of assembly operations, jamming can occur due to uncertainties that needs be compensated.
Learning human-inspired skills through imitation from a assembly process.
An adaptive control model is developed using recursive neural networks (RNNs) with long short-term memory (LSTM) architecture, which can remember control actions and further improve them. The control model is implemented as cartesian force control with forward dynamics. Initial and safe imitation learning is provided by human teleoperation in the simulation. A robust Sim2Real transfer enables easy and fast deployment in real applications.
The adaptive controllers are available as open source code, which enables a rapid transfer to production.