Neurorobotic reinforcement learning for domains with parametrical uncertainty

Area of Science:

• Robotics
• Artificial Intelligence
• Neuroscience

Background:

• Neuromorphic hardware offers low energy consumption and latency for robot control.
• Brain-inspired learning strategies are crucial for advancing robotic capabilities.
• Integrating neuromorphic systems in simulation is key to improving robot control.

Purpose of the Study:

• To implement spiking reinforcement learning for robotic arm control using neuromorphic hardware.
• To address the sim-to-real gap in robotic control through randomized simulation parameters.
• To demonstrate the first neuromorphic implementation of the peg-in-hole task with hardware-in-the-loop.

Main Methods:

• Utilized the neurorobotics platform (NRP) simulation framework.
• Implemented spiking reinforcement learning with force-torque feedback for a peg-in-hole task.
• Integrated the neuromorphic Loihi chip for hardware-in-the-loop control and utilized scripted accelerated training.

Main Results:

• Successfully controlled a robotic arm using neuromorphic hardware in a simulated peg-in-hole task.
• Developed robust policies by training in randomized simulation environments.
• Achieved effective control despite real-world parameter variations, bridging the sim-to-real gap.

Conclusions:

• Neuromorphic hardware and spiking reinforcement learning provide a viable solution for robust robot control.
• The study demonstrates a significant advancement in neuromorphic robotics, particularly for tasks requiring precise manipulation.
• This work paves the way for more adaptable and efficient robotic systems in complex environments.