Simulation and Sim-to-real
Simulation and Sim-to-real#
{"description": "Training in simulation and sim-to-real transfer in Duckietown.", "keywords": "reinforcement learning, sim-to-real, duckietown, machine learning, ML, AI, embedded AI"}
There are several reasons why training RL agents on real robot hardware is challenging. Training in a simulator and then applying the resulting model in the real world can be a good strategy to overcome some of these challenges. This transfer process is referred to as sim-to-real transfer.
In general, there are two classes of approaches. The first is to explicitly model the discrepancies between the simulated and real worlds. The second is to use the simulator and RL training scheme to learn a policy that is inherently robust to these differences.