Abstract
Robotic technology is integrating into more aspects of our society. The ability to perform
multiple different functions in various environments is an important research challenge. Recent
developments in multi-task learning in Artificial Intelligence (AI) and machine learning could
enable a new generation of robotic technology. This thesis aims to explore techniques which
combine AI with robotics to enable generalization across a range of environments.
As the level of abstraction possible in deep learning has grown a very promising avenue of
research has become feasible: hybrid systems. In these approaches, an AI system is used to
undertake high-level planning and strategisation, while the execution of the plans is undertaken
by classical non-learned approaches. This enables a mutual compensation of AI and traditional
robotic systems, as the AI agents excel at generalisation for high-level problems and longterm
goal-oriented planning, while the robotics techniques specialise in precise and repeatable operations [24].
This thesis attempts to build upon this and further explore the multi-task-enabled AI agent and its
potential in robotic tasks.
First, general-purpose robots need to be able to work in multiple different environments. Even
when performing similar tasks, different behaviour should be deployed to best fit the current
environment. In this thesis, we propose a new approach to navigation, where it is treated
as a multi-task learning problem. This enables the robot to learn to behave differently in
visual navigation tasks for different environments while also learning shared expertise across
environments. We evaluate our approach in both simulated environments as well as real-world
data.
Following this, we introduce a new perspective for learning transferable content in multi-task
imitation learning. Humans are able to transfer skills and knowledge to new tasks. If we can
cycle to work and drive to the store, we can also cycle to the store and drive to work. We take
inspiration from this and hypothesize the latent memory of a policy network can be disentangled
into two partitions. These contain either the knowledge of the environmental context for the task
or the generalizable skill needed to solve the task. This allows improved training efficiency and
better generalization over previously unseen combinations of skills in the same environment, and
the same task in unseen environments. We used the proposed approach to train a disentangled
agent for two different multi-task IL environments. In both cases, we outperformed the SOTA
by 30% in task success rate. We also demonstrated this for navigation on a real robot, including
the visual navigation task of the previous chapter. This demonstration was also ported to a real
robot.
This ability to generalize to previously unseen tasks with little to no supervision is a key
challenge in modern machine learning research. Researchers often rely on reinforcement and
imitation learning to provide online adaptation to new tasks, through trial and error learning.
However, this can be challenging for complex tasks which require many timesteps or large
numbers of subtasks to complete. These “long horizon” tasks suffer from sample inefficiency
and can require extremely long training times before the agent can learn to perform the necessary
long-term planning. Therefore, we finally introduce CASE which attempts to address these
issues by utilising adaptive “near future” subgoals. These subgoals are re-calculated at each step
using compositional arithmetic in a learned latent representation space similar to the previous
chapter. In addition to improving learning efficiency for standard long-term tasks, this approach
also makes it possible to perform one-shot generalization to previously unseen tasks, given only
a single reference trajectory for the task in a different environment.
When we combine the works in this thesis, these contributions provide insights for a better
multitask generalization framework for long and complex robotic planning. The contributions centre
on multi-task learning and crossover to robotic navigation, reinforcement/imitation learning,
and generalization. The experiments in this thesis demonstrate that an AI-controlled robot can
perform simple tasks as well as a non-AI robot, while also handling more difficult tasks. This
thesis presents a general direction for robotic AI capable of handling multiple complex tasks and
sequential generalization. This enables better autonomy in both AI and robotics in the future.