Simulation In The Loop : Real-Time Robot Action Success Verification with a Digital Twin

Sammanfattning

This thesis investigates how robotic task execution can be improved by leveraging a digital twin to perform action success verification, in a paradigm referred to as Simulation-In-The-Loop (SIL). Specifically, this work focuses on using SIL for environment resetting, a process that returns a robot– environment system to a known initial state, which when implemented successfully supports continuous data collection and iterative learning. Using a digital twin constructed by the author of this thesis, SIL enables action validation through simulated perturbations prior to real-world deployment. By integrating cloud-based simulations, parallel testing of candidate actions is conducted efficiently and a standardized toolkit provides consistent evaluation metrics. Experiments with a 3-DoF robot arm revealed that SIL can substantially improve task success, with the observed task success rate increasing by 113% compared to a baseline approach, which performed task sequences without verifying actions in simulation. However, the relatively simple environment-resetting task in the study and the reliance on a specific hardware platform limit direct generalization to more complex or dynamic scenarios. Discrepancies between the simulated and real-world conditions (reality gap) further restricted the transferability of simulation-validated actions. These findings indicate that SIL may offer a promising pathway to enhance sample efficiency and reliability in robotic tasks, especially when combined with scalable, cloud-based infrastructure. Future work includes expanding the method to more diverse tasks, improving simulation fidelity, allowing users to customize position definitions, and transferring the pipeline to a fully cloud-based infrastructure. Although additional investigations are needed to validate SIL in complex settings, this thesis demonstrates how simulation-driven action perturbation can contribute to the design and deployment of more robust autonomous systems.