Transferring robot manipulation policies from simulation to real-world
Lead Research Organisation:
Imperial College London
Department Name: Computing
Abstract
Robots are increasingly becoming part of our lives, but despite their impressive repertoire of tasks, many of them will fail to adapt when presented to new and unfamiliar environments. Before robots can realise their full potential in everyday life, they need the ability to manipulate the changing world around them. Recent trends to solve this problem have seen a shift to end-to-end solutions using deep reinforcement learning policies from visual input. However, these approaches are run on real-world robotic platforms which often require human interaction, expensive equipment, and long training times. We aim to explore the feasibility of using life-like simulations to train robot manipulations tasks that are then directly map to real-world hardware
Organisations
People |
ORCID iD |
Andrew Davison (Primary Supervisor) | |
Stephen James (Student) |
Publications
Bonardi A
(2020)
Learning One-Shot Imitation From Humans Without Humans
in IEEE Robotics and Automation Letters
James S
(2020)
RLBench: The Robot Learning Benchmark & Learning Environment
in IEEE Robotics and Automation Letters
Studentship Projects
Project Reference | Relationship | Related To | Start | End | Student Name |
---|---|---|---|---|---|
EP/N509486/1 | 30/09/2016 | 30/03/2022 | |||
1793137 | Studentship | EP/N509486/1 | 30/09/2016 | 30/03/2020 | Stephen James |
Description | We have discovered that instead of explicitly modelling each stage of the robot manipulation pipeline, we are instead able to learn how to do tasks in a fully end-to-end way. Moreover, we have discovered that it is possible to learn such skills in simulation and then transfer this knowledge from simulation to the real-world. |
Exploitation Route | Advances in machine learning and robotics will enable household robotics. |
Sectors | Manufacturing including Industrial Biotechology Retail Other |