Research in the Cognitive Machines Group is based on an interplay of simulations and real-world robotics with the goal of developing language processing systems with physically-grounded semantics. Simulations provide a fast means for developing early ideas. These ideas are cached out in the real world using robots with sensors and actuators. This methodology provides the best of both worlds. On one hand, virtual worlds are easy to build and provide a clean development environment. The simulations serve a critical role as scaffolding for their real-world counterparts. Inevitably, however, simulations capture only part of the actual complexity of the world. Building physically embodied systems forces us to confront those complexities head-on. Ultimately, we are committed to implementing physically-instantiated systems as a safeguard against relying on potentially false human intuitions about how the world works.