Reliable reinforcement learning for complex robot control under real-world constraints

PhD progress review 3 presentation:

 

Reliable reinforcement learning for complex robot control under real-world constraints

 

Speaker: Humphrey Munn

Abstract:
Reinforcement learning can enable robots to complete complex tasks in
simulation, but making those behaviours work reliably on real hardware
is harder. Real robots must operate under imperfect models, changing
conditions, competing task objectives, and hard physical constraints on
stability, safety, compliance, and energy. In this PhD seminar, I will
present the main work from my thesis on reliable reinforcement learning
for dynamic robot control. The talk will cover four projects that
approach this problem from different angles: teaching legged robots to
perform dynamic throwing, where precision depends on complex interaction
between the lower body and upper body while maintaining realistic
stability; improving state-of-the-art reinforcement learning algorithms
when robot tasks contain many conflicting objectives; monitoring learned
policies during deployment to detect and diagnose failures; and adapting
pre-trained policies when the robot’s physical conditions change, such
as when carrying payloads or losing actuator authority. Across these
studies, the central question is how to make learned controllers
dependable once they leave the simulator. I will discuss what worked,
what remained difficult, and how successful real-world deployment
depends on multiple connected components, from training and optimisation
to modelling fidelity, monitoring, and adaptation, rather than policy
performance alone.