Virtual AI training platform for robot learning

 

AUDI, TUM

The demonstrator uses Reinforcement Learning (RL) to address the challenge of bringing autonomy in industrial robotic manipulation by implementing the first necessary steps. The biggest bottleneck of learning an optimal policy in robotics through RL is the need to provide real-world environments in which a robot can actively operate and safely explore different policies for an extended period. Therefore, advanced simulations are used to virtually train the policy network by providing multitudes of realistic synthetic data. Based on the current state-of-the-art methodologies in deep RL, virtual training in simulations is the best feasible solution for realising autonomous artificial agents embodied in robots.

Beyond state-of-the-art developments and impacts

The RL in robot learning is still a hot unsolved topic for researchers worldwide. No generic standard requirements exist yet for benchmarking or verifying if the trained deep model is functioning correctly. Novel functional requirements are identified to demonstrate the framework's applicability, compatibility, and scalability for learning various industry-relevant robotic manipulation tasks.