Training model-free reinforcement learning algorithms currently requires a very large amount of agent interaction to solve challenging tasks, often far more interaction than would be practical on a robot in real time. In addition, photorealistic simulations of specific environments can be hard to come by. For these reasons, we propose learning as much as possible directly from real recorded data.

This project page presents visualisations of the results presented in the paper, and provides the code and data required to reproduce the results from “Learning Deployable Navigation Policies at Kilometer Scale from a Single Traversal”, and apply the approach to other datasets and robots.

CampusMap

Entire Dataset

This animation shows one frame for each of the 2099 discrete locations in the dataset.

EntireDatasetGif

Experiment Videos

The following animations show one in every 60 frames of the deployment trajectories. This equates to approximately one frame per meter at maximum driving speed. Below each animation is a plot of the trajectory taken at deployment time compared against the optimal trajectory and the trajectory the agent executes on the training data.

Trajectory 1 (successful)

Trj1
Trj1 Trj1 Trj1

Trajectory 2 (successful)

Trj2
Trj1 Trj1 Trj1

Trajectory 3 (successful)

Trj3
Trj1 Trj1 Trj1

Trajectory 4 (successful)

Trj4
Trj1 Trj1 Trj1

Trajectory 5 (successful)

Trj5
Trj1 Trj1 Trj1

Trajectory 6 (failed)

Trj6
Trj1 Trj1 Trj1

Trajectory 7 (successful)

Trj7
Trj1 Trj1 Trj1

Trajectory 8 (successful)

Trj8
Trj1 Trj1 Trj1

Trajectory 9 (successful)

Trj9
Trj1 Trj1 Trj1

Trajectory 10 (successful)

Trj10
Trj1 Trj1 Trj1

Trajectory 11 (successful)

Trj11
Trj1 Trj1 Trj1

Video of Trained Agent

The following video shows the trained agent navigating in recorded data with a gradually increasing curriculum.