Fraunhofer Institute for Material Flow and Logistics IML
Simulation-based AI: Key Technology for the Development of Autonomous Logistics Robots
The use of autonomous systems in logistics and other domains is rapidly increasing. This implies an icreas of the complexity of these systems, caused due to varrious demands such as higher speeds, able to handle new tasks, or the cooperation between robots and humans. This makes the development of autonomous systems more challenging and at the same time more costly and time-consuming. At the same time, extensive data is required to train AI solutions, and the broad application of the systems requires the testing of many scenarios.
Simulation-based AI can be a key technology for the development of autonomous robotic logistics systems. For this purpose, we use advanced simulation tools from NVIDIA Omniverse, including Isaac Sim and Isaac Lab. AI methods are used to create the simulation models, while at the same time various AI models can be trained with the simulation data. Overall, such (self-)learning models of real systems offer numerous advantages, e.g.,:
1. Parallelized hardware and software development.
2. Resource-efficient training of new algorithms in realistic virtual environments.
3. Use of synthetic sensor data for training computer vision algorithms.
4. Pre-testing of hardware modifications in the virtual digital world, reducing the number of prototypes needed and thus the consumption of resources.
Overall, simulation-based AI enables accelerated planning and development of logistical robot systems.
The basis for simulation-based AI is a holistic modeling of the robot systems including actuators, sensors, and communication interfaces. By using design data and incorporating real data from the system, the robot's physical behavior can be modeled accurately. We record the real data and motion data of mobile robots in our own infrastructure, such as the PACE Lab. New rendering technologies in GPU-based simulation tools enable the mapping of complex sensors such as multi-layer laser sensors and high-resolution cameras. By replicating the system's communication interfaces, a holistic modeling is achieved and the difference between simulated and real robots is minimized piece by piece. In the best case, there is no difference between the robot and the model.
We have demonstrated the potential of simulation-based AI in various works with our robotic platforms developed at the IML. The Isaac Sim simulation models from O3dyn and evoBOT are available as open source references. High-ranking scientific publications and various lectures, including at ICRA, IROS NVIDIA GTC and ROSCon, underline this potential.
