We present FORGE, a method that enables sim-to-real transfer of contact-rich manipulation policies in the presence of significant pose uncertainty. FORGE combines a force-threshold mechanism with a dynamics randomization scheme during policy learning in simulation, to enable the robust transfer of the learned policies to the real robot. At deployment, FORGE policies, conditioned on a maximum allowable force threshold, adaptively perform contact-rich tasks while respecting the specified force threshold, regardless of the controller gains. Additionally, FORGE autonomously predicts a termination action, once the task has succeeded. We demonstrate that FORGE can be used to learn a variety of robust contact-rich policies, enabling multi-stage assembly of a planetary gear system, which requires success across three assembly tasks: nut-threading, insertion, and gear meshing.
This research was sponsored in part by Army Research Labs under the A2I2 program and by NVIDIA. Their support is gratefully appreciated.
