Reproduce “Safety-Driven Deep Reinforcement Learning Framework for Cobots: A Sim2Real Approach” on OpenResearch

About this paper

Ammar N. Abbas, Shakra Mehak, Georgios C. Chasparis, John D. Kelleher, Michael Guilfoyle, Maria Chiara Leva, Aswin K Ramasubramanian

A Safety-Driven Deep Reinforcement Learning (SD-DRL) framework integrates industrial safety standards directly into a robot's reward function, enabling cobots to learn inherently safe behaviors. This approach demonstrated improved task success rates by an average of 16.5% in simulations and 2.5% in real-world tests without safety violations, achieving Safety Integrity Level 2 compliance.

Key insights

• Integrating industrial safety standards directly into the DRL reward function enables robots to learn inherently safer behaviors without significantly compromising task success.
• DRL-driven robotic systems can undergo functional safety assessments, achieving quantifiable Safety Integrity Levels (SILs) compliant with standards like ISO/IEC 61508.
• The fidelity of the simulation's physics engine can impact the Sim2Real transfer of physics-based safety parameters, indicating a need for more accurate simulation modeling to fully bridge the gap.

Summary by alphaXiv.