Experimental study demonstrates sim-to-real transition control in a mini rotary double inverted pendulum, highlighting the viability of physics-based reinforcement learning models.
Key Points
To develop and evaluate a sim-to-real reinforcement learning framework for controlling transitions between multiple equilibrium states in an underactuated mini rotary double inverted pendulum.
Engineered a direct-drive rotary double inverted pendulum system with torque-controlled actuation and constructed a high-fidelity MuJoCo model incorporating CAD and experimentally identified dynamic parameters.
Trained multi-equilibrium transition policies using the Truncated Quantile Critics (TQC) deep reinforcement learning algorithm.
Directly transferred trained simulation policies to the physical hardware and assessed performance across successive equilibrium transitions without resetting the system.
Simulation-trained policies achieved zero-shot sim-to-real transfer, successfully executing intended transitions between multiple equilibrium points on the physical hardware.
The direct-drive torque control and identified parameters enabled continuous, back-to-back state transitions without requiring manual resets between trials.