The next IPAB Workshop will take place on 05/12/2019 at 12:45 in room G.03 (pastries will be served). Todor Davchev, Daniel Gordon will be speaking.
Speaker: Todor Davchev
Title: Robust Skill Acquisition with Residual Policy Learning
Abstract: Dealing with contacts and friction is a common requirement for modern manufacturing. However, relying on a control engineer to manually tune every deployed conventional feedback controller is often expensive and limited to a specific task. Reinforcement learning (RL) methods, on the other hand, have been shown to successfully learn continuous robot controllers for physical skill acquisition tasks on the often expensive cost of hyper-parameter tuning and data collection. In this work, we propose a framework that can acquire the skill of insertion in the context of a series of nonlinear physical tasks solved by a robot arm. We formulate our solution as a dynamic movement primitive that relies on behavioural cloning to extract an initial policy. We show that combining data aggregation techniques for behavioural cloning with a residual policy learning component solved with RL can successfully learn to complete a task in one hour. The final learned policy is then defined as a combination of both control signals. We demonstrate this approach by training an agent to successfully perform real-world peg and plug to socket insertions.
Speaker: Daniel Gordon
Title: Human-in-the-loop Optimisation of Exoskeleton Assistance Patterns
Abstract: Over the past few decades, many assistive robotic devices have been developed. Despite advancements in design and control algorithms, the problem of assisting locomotion remains challenging. Human walking strategies are unique and complex, and assistance strategies based on the dynamics of unassisted locomotion typically offer only modest reductions to the metabolic cost of walking. Recently, human-in-the-loop (HIL) methodologies have been used to identify assistive strategies which offer significant improvements to energy savings. However, current implementations suffer from long measurement times, necessitating the use of low-dimensional control parameterisations, and possibly requiring multi-day collection protocols to avoid subject fatigue. We present a HIL methodology which optimises the assistive torques provided by a powered hip exoskeleton. Using musculoskeletal modelling, we are able to evaluate simulated metabolic rate online, which reduces measurement times compared to previous methods. Our framework could be used to enable shorter HIL protocols or explore more complex control parameterisations.