Neural real-time planning for reactive industrial robots
Neuromorphic systems are known for fast, highly parallel and energy efficient computation. ln addition, their ability to learn and adapt makes them very flexible and tolerant to noise and hardware failure. The recent emergence of dedicated, neuromorphic computation hardware and novel sensors opens the opportunity tcomake these advantages accessible to industrial robots. The NeuroReact project demonstrates this by combining neuromorphic hardware and findings from neurorobotics to create a realtime capable, reactive planning system.
Learning only plays a subordinate role. The inverse kinematics are implicitly coded by the transformatior trom Cartesian space to configuration space. The configuration space embodies all possible combinations of the robot's joint angles. This is not done by learning but by sampling in simulation. A demonstrator is planned, with a UR5 or URlO performing a picking task with collision avoidance.
Collaboration with HBP
The techniques and practices of NeuroReact are going to be published as an open source project and thus also be accessible for the HBP.
Furthermore NeuroReact uses the neuromorphic hardware SpiNNaker in an industry-related context. Spiking Neural Networks (SNN) and also neuromorphic hardware, are, despite their great potential, still merely used academically. A motivation of the project is to change this and open these systems up to the public. This might be especially interesting for HBP, since it has to finance itself after the European funding period of ten years. Hence it would be beneficial for the HBP if core parts like SpiNNaker and the NRP increase their popularity outside of the academic community.
Time frame: 2017-2020
Origin: Spontaneous Application
Funding: Baden-Wuertemmberg Stiftung