This repository contains an implementation of the StreetLearn C++ engine and Python environment for training navigation agents in real-world photographic street environments, as well as code for implementing the agents used in [1] "Learning to Navigate in Cities Without a Map" (NeurIPS 2018). This environment was also used in two follow-up papers: [2] "Cross-View Policy Learning for Street Navigation" (ICCV 2019) and [3] "Learning to follow directions in Street View" (AAAI 2020), as well as in technical report [4] "The StreetLearn Environment and Dataset". The StreetLearn environment relies on panorama images from Google Street View and provides an interface for moving a first-person view agent inside the Street View graph. This is not an officially supported Google product. Please cite papers [1] and [4] if you use the code from this repository in your work.

Our papers [1], [2] and [3] also provide a detailed description of how to train and implement navigation agents in the StreetLearn environment by using a TensorFlow implementation of "Importance Weighted Actor-Learner Architectures", published in Espeholt, Soyer, Munos et al. (2018) "IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures"). The generic agent and trainer code have been published by Lasse Espeholt under an Apache license at: https://github.com/deepmind/scalable_agent.

@inproceedings{mirowski2018learning,
  title={Learning to Navigate in Cities Without a Map},
  author={Mirowski, Piotr and Grimes, Matthew Koichi and Malinowski, Mateusz and Hermann, Karl Moritz and Anderson, Keith and Teplyashin, Denis and Simonyan, Karen and Kavukcuoglu, Koray and Zisserman, Andrew and Hadsell, Raia},
  booktitle={Neural Information Processing Systems (NeurIPS)},
  year={2018}
}

@article{mirowski2019streetlearn,
  title={The StreetLearn Environment and Dataset},
  author={Mirowski, Piotr and Banki-Horvath, Andras and Anderson, Keith and Teplyashin, Denis and Hermann, Karl Moritz and Malinowski, Mateusz and Grimes, Matthew Koichi and Simonyan, Karen and Kavukcuoglu, Koray and Zisserman, Andrew and others},
  journal={arXiv preprint arXiv:1903.01292},
  year={2019}
}

This environment code contains:

• streetlearn/engine Our C++ StreetLearn engine for loading, caching and serving Google Street View panoramas by projecting them from a equirectangular representation to first-person projected view at a given yaw, pitch and field of view, and for handling navigation (moving from one panorama to another) depending on the city street graph and the current orientation.
• streetlearn/proto The message protocol buffer used to store panoramas and street graph.
• streetlearn/python/environment A Python-based interface for calling the StreetLearn environment with custom action spaces. Within the Python StreetLearn interface, several games are defined in individual files whose names end with game.py. A second interface, called BatchedStreetLearn, can be used to instantiate multiple StreetLearn environments that share the same action specs, observation specs, and panorama cache, and return observations in batched format.
• streetlearn/python/ui A simple interactive human_agent and an oracle_agent and instruction_following_oracle_agent for courier and instruction-following tasks respectively; all agents are implemented in Python using pygame and instantiate the StreetLearn environment on the requested map, along with a simple user interface. The interactive human_agent enables a user to play various games. The oracle_agent and instruction_following_oracle_agent are similar to the human agent and automatically navigate towards the goal (courier game) or towards the goal via waypoints, following instructions (instruction-following game) and they report oracle performance on these tasks. A batched version of th oracle agent can be started using batched_oracle_agent.

Bazel is the official build system for StreetLearn. The build has only been tested running on Ubuntu 18.04.

sudo apt-get install autoconf automake libtool curl make g++ unzip virtualenv python-virtualenv cmake subversion pkg-config libpython-dev libcairo2-dev libboost-all-dev python-pip libssl-dev
pip install setuptools
pip install pyparsing

For detailed information see: https://github.com/protocolbuffers/protobuf/blob/master/src/README.md

git clone https://github.com/protocolbuffers/protobuf.git
cd protobuf
git submodule update --init --recursive
./autogen.sh
./configure
make -j7
sudo make install
sudo ldconfig
cd python
python setup.py build
sudo python setup.py install
cd ../..

git clone https://github.com/google/clif.git
cd clif
./INSTALL.sh
cd ..

wget https://github.com/opencv/opencv/archive/2.4.13.6.zip
unzip 2.4.13.6.zip
cd opencv-2.4.13.6
mkdir build
cd build
cmake -D CMAKE_BUILD_TYPE=Release -D CMAKE_INSTALL_PREFIX=/usr/local ..
make -j7
sudo make install
sudo ldconfig
cd ../..

pip install six
pip install absl-py
pip install inflection
pip install wrapt
pip install numpy
pip install dm-sonnet
pip install tensorflow-gpu
pip install pygame

This page describes how to install the Bazel build and test tool on your machine. We currently support Bazel versions up to 0.24.0, whose installation instructions are listed on this page under section Using the binary installer (copy-pasted below):

wget https://github.com/bazelbuild/bazel/releases/download/0.24.0/bazel-0.24.0-installer-linux-x86_64.sh
chmod +x bazel-0.24.0-installer-linux-x86_64.sh
./bazel-0.24.0-installer-linux-x86_64.sh  --user
export PATH="$PATH:$HOME/bin"

Clone this repository and install Scalable Agent:

git clone https://github.com/deepmind/streetlearn.git
cd streetlearn
sh get_scalable_agent.sh

To build the StreetLearn engine only:

export CLIF_PATH=$HOME/opt
bazel build streetlearn:streetlearn_engine_py

To build the human agent and the oracle agent in the StreetLearn environment, with all the dependencies:

export CLIF_PATH=$HOME/opt
bazel build streetlearn/python/ui:all

To run the human agent using one of the StreetLearn datasets downloaded and stored at dataset_path (note that dataset_path needs to be absolute, not relative):

bazel run streetlearn/python/ui:human_agent -- --dataset_path=<dataset_path>

For help with the options of the human_agent:

bazel run streetlearn/python/ui:human_agent -- --help

Similarly, to run the oracle agent on the courier game:

bazel run streetlearn/python/ui:oracle_agent -- --dataset_path=<dataset_path>

The human agent and the oracle agent show a view_image (on top) and a graph_image (on bottom).

Note: you need to add the following line to your .bashrc script to avoid specifying the CLIF path each time you open a new terminal:

export CLIF_PATH=$HOME/opt

• Rotate left or right in the panorama, by a specified angle (change the yaw of the agent). In the human_agent, press a or d.
• Rotate up or down in the panorama, by a specified angle (change the pitch of the agent). In the human_agent, press w or s.
• Move from current panorama A forward to another panorama B if the current bearing of the agent from A to B is within a tolerance angle of 30 degrees. In the human_agent, press space.
• Zoom in and out in the panorama. In the human_agent, press i or o.

Additional keys for the human_agent are escape and p (to print the current view as a bitmap image).

For training RL agents, action spaces are discretized using integers. For instance, in our paper, we used 5 actions: (move forward, turn left by 22.5 deg, turn left by 67.5 deg, turn right by 22.5 deg, turn right by 67.5 deg).

Along the bottom of the view_image is the navigation bar which displays a small circle in any direction in which travel is possible:

• When within the centre range, they will turn green meaning the user can move in this direction.
• When they are out of this range, they will turn red meaning this is inaccessible.
• When more than one dots are within the centre range, all except the most central will turn orange, meaning that there are multiple (forward) directions available.

The graph is constructed by breadth first search to the depth specified by the graph depth flags. At the maximum depth the graph will suddenly stop, generally in the middle of a street. Because we are trying to train agents to recognize streets as navigable, and in order not to confuse the agents, red stop signs are shown from two panoramas away from any terminal node in the graph.

You can request the StreetLearn dataset on the StreetLearn project website. The following datasets are currently distributed:

• 56k Manhattan panoramas, used [1], [2], [3] and [4]: ** manhattan_highres (size 1632 x 408) ** manhattan_lowres (size 408 x 204)
• 58k Pittsburgh panoramas, used in [2], [3] and [4]: ** pittsburgh_highres (size 1632 x 408) ** pittsburgh_lowres (size 408 x 204)
• 29k Manhattan panoramas used in [5] "TOUCHDOWN: Natural Language Navigation and Spatial Reasoning in Visual Street Environments" (Chen, Suhr, Misra et al, ICCV 2019), with accompanying code at https://github.com/lil-lab/touchdown: ** touchdown_manhattan_highres (size 3000 x 1500) ** touchdown_manhattan_lowres (downsampled to 500 x 250)

The downsampled version of the panoramas can be used when the RGB inputs are small (e.g., 84 x 84), to save on panorama image loading and projection time.

The Python StreetLearn environment follows the specifications from OpenAI Gym. The call to function step(action) returns:

• observation (tuple of observations requested at construction),
• reward (a float with the current reward of the agent),
• done (boolean indicating whether the episode has ended)
• and info (a dictionary of environment state variables). After creating the environment, it is initialised by calling function reset(). If the flag auto_reset is set to True at construction, reset() will be called automatically every time that an episode ends.

Default environment settings are stored in streetlearn/python/default_config.py.

• seed: Random seed.
• width: Width of the streetview image.
• height: Height of the streetview image.
• graph_width: Width of the map graph image.
• graph_height: Height of the map graph image.
• status_height: Status bar height in pixels.
• field_of_view: Horizontal field of view, in degrees.
• min_graph_depth: Min bound on BFS depth for panos.
• max_graph_depth: Max bound on BFS depth for panos.
• max_cache_size: Pano cache size.
• bbox_lat_min: Minimum value for normalizing the target latitude.
• bbox_lat_max: Maximum value for normalizing the target latitude.
• bbox_lng_min: Minimum value for normalizing the target longitude.
• bbox_lng_max: Maximum value for normalizing the target longitude.
• min_radius_meters: Minimum distance from goal at which reward shaping starts in the courier game.
• max_radius_meters: Maximum distance from goal at which reward shaping starts in the courier game.
• timestamp_start_curriculum: Integer timestamp (UNIX time) when curriculum learning starts, used in the curriculum courier game.
• hours_curriculum_part_1: Number of hours for the first part of curriculum training (goal location within minimum distance), used in the curriculum courier game.
• hours_curriculum_part_2: Number of hours for the second part of curriculum training (goal location annealed further away), used in the curriculum courier game.
• min_goal_distance_curriculum: Distance in meters of the goal location at the beginning of curriculum learning, used in the curriculum courier game.
• max_goal_distance_curriculum: Distance in meters of the goal location at the beginning of curriculum learning, used in the curriculum courier game.
• instruction_curriculum_type: Type of curriculum learning, used in the instruction following games.
• frame_cap: Episode frame cap.
• full_graph: Boolean indicating whether to build the entire graph upon episode start.
• sample_graph_depth: Boolean indicating whether to sample graph depth between min_graph_depth and max_graph_depth.
• start_pano: The pano ID string to start from. The graph will be build out from this point.
• graph_zoom: Initial graph zoom. Valid between 1 and 32.
• graph_black_on_white: Show the graph as black on white. Default value: false (shows the graph as white on black).
• show_shortest_path: Boolean indicator asking whether the shortest path to the goal shall be shown on the graph.
• calculate_ground_truth: Boolean indicator asking whether the ground truth direction to the goal should be calculated during the game (useful for oracle agents, visualisation and for imitation learning).
• neighbor_resolution: Used to calculate a binary occupancy vector of neighbors to the current pano.
• color_for_touched_pano: RGB color for the panos touched by the agent.
• color_for_observer: RGB color for the observer.
• color_for_coin: RGB color for the panos containing coins.
• color_for_goal: RGB color for the goal pano.
• color_for_shortest_path: RGB color for panos on the shortest path to the goal.
• color_for_waypoint: RGB color for a waypoint pano.
• observations: Array containing one or more names of the observations requested from the environment: ['view_image', 'graph_image', 'yaw', 'pitch', 'metadata', 'target_metadata', 'latlng', 'target_latlng', 'latlng_label', 'target_latlng_label', 'yaw_label', 'neighbors', 'thumbnails', 'instructions', 'ground_truth_direction']
• reward_per_coin: Coin reward for coin game.
• reward_at_waypoint: Waypoint reward for the instruction-following games.
• reward_at_goal: Goal reward for the instruction-following games.
• proportion_of_panos_with_coins: The proportion of panos with coins.
• game_name: Game name, can be: 'coin_game', 'exploration_game', 'courier_game', 'curriculum_courier_game', 'goal_instruction_game', 'incremental_instruction_game' and 'step_by_step_instruction_game'.
• action_spec: Either of 'streetlearn_default', 'streetlearn_fast_rotate', 'streetlearn_tilt'
• rotation_speed: Rotation speed in degrees. Used to create the action spec.
• auto_reset: Boolean indicator whether games are reset automatically when the max number of frames is achieved.

The following observations can be returned by the agent:

• view_image: RGB image for the first-person view image returned from the environment and seen by the agent,
• graph_image: RGB image for the top-down street graph image, usually not seen by the agent,
• yaw: Scalar value of the yaw angle of the agent, in degrees (zero corresponds to North),
• pitch: Scalar value of the pitch angle of the agent, in degrees (zero corresponds to horizontal),
• metadata: Message protocol buffer of type Pano with the metadata of the current panorama,
• target_metadata: Message protocol buffer of type Pano with the metadata of the target/goal panorama,
• latlng: Tuple of lat/lng scalar values for the current position of the agent,
• target_latlng: Tuple of lat/lng scalar values for the target/goal position,
• latlng_label: Integer discretized value of the current agent position using 1024 bins (32 bins for latitude and 32 bins for longitude),
• target_latlng_label: Integer discretized value of the target position using 1024 bins (32 bins for latitude and 32 bins for longitude),
• yaw_label: Integer discretized value of the agent yaw using 16 bins,
• neighbors: Vector of immediate neighbor egocentric traversability grid around the agent, with 16 bins for the directions around the agent and bin 0 corresponding to the traversability straight ahead of the agent.
• thumbnails: Array of n+1 RGB images for the first-person view image returned from the environment, that should be seen by the agent at specific waypoints and goal locations when playing the instruction-following game with n instructions,
• instructions: List of n string instructions for the agent at specific waypoints and goal locations when playing the instruction-following game with n instructions,
• ground_truth_direction: Scalar value of the relative ground truth direction to be taken by the agent in order to follow a shortest path to the next goal or waypoint. This observation should be requested only for agents trained using imitation learning.

The following games are available in the StreetLearn environment:

• coin_game: invisible coins scattered throughout the map, yielding a reward of 1 for each. Once picked up, these rewards do not reappear until the end of the episode.
• courier_game: the agent is given a goal destination, specified as lat/long pairs. Once the goal is reached (with 100m tolerance), a new goal is sampled, until the end of the episode. Rewards at a goal are proportional to the number of panoramas on the shortest path from the agent's position when it gets the new goal assignment to that goal position. Additional reward shaping consists in early rewards when the agent gets within a range of 200m of the goal. Additional coins can also be scattered throughout the environment. The proportion of coins, the goal radius and the early reward radius are parametrizable.
• curriculum_courier_game: same as the courier game, but with a curriculum on the difficulty of the task (maximum straight-line distance from the agent's position to the goal when it is assigned).
• goal_instruction_game and its variations incremental_instruction_game and step_by_step_instruction_game use navigation instructions to direct agents to a goal. Agents are provided with a list of instructions as well as thumbnails that guide the agent from its starting position to the goal location. In step_by_step_instruction_game, agents are provided one instruction and two thumbnails at a time, in the other game variants the whole list is available throughout the whole game. Reward is granted upon reaching the goal location (all variants), as well as when hitting individual waypoints (incremental_instruction_game and step_by_step_instruction_game only). During training various curriculum strategies are available to the agents, and reward shaping can be employed to provide fractional rewards when the agent gets within a range of 50m of a waypoint or goal.

The Abseil C++ library is licensed under the terms of the Apache license. See LICENSE for more information.

This is not an official Google product.