The Traffic Manager, TM for short, is the module in charge of controlling vehicles inside the simulation. It is built on top of the CARLA API in C++. Its goal is to populate the simulation with realistic urban traffic conditions. Users can customize some behaviours, for example to set specific learning circumstances. Every TM controls vehicles registered to it by setting autopilot to true, and accounts for the rest by considering them unregistered.
The TM is built on the client-side of the CARLA architecture. It replaces the server-side autopilot. The execution flow is divided in __stages__ with independent operations and goals. This facilitates the development of phase-related functionalities and data structures, while also improving computational efficiency. Each stage runs on a different thread. Communication with the rest is managed through synchronous messaging between stages. Information flows only in one direction.
Users have some control over the traffic flow by setting parameters that allow, force or encourage specific behaviours. Users can change the traffic behaviour as they prefer, both online and offline. For example they could allow a car to ignore the speed limits or force a lane change. Being able to play around with behaviours is a must when trying to simulate reality. It is necessary to train driving systems under specific and atypical circumstances.
The previous diagram is a summary of the internal architecture of the Traffic Manager. The inner structure of the TM can be easily translated to code, and each relevant component has its equivalent in the C++ code (.cpp files) inside `LibCarla/source/carla/trafficmanager`. The functions and relations of these components are explained in the detail in the following sections. A simplified overview of the logic is as follows:
First of all, the [ALSM](#alsm) (Agent Lifecycle & State Management) scans the world to keep track of all the vehicles and walkers present and to clean up entries for those that no longer exist. All the data is retrieved from the server, and then passed to the [stages](#stages). In such way, calls to the server are isolated in the ALSM, and this information can be easily accessible onwards. The [vehicle registry](#vehicle-registry) contains an array with the registered vehicles, and a list with the rest of vehicles and pedestrians. The [simulation state](#simulation-state) stores in cache the position and velocity and some additional information of all the cars and walkers.
The main goal of the TM is to generate viable commands for all the vehicles in the [vehicle registry](#vehicle-registry), according to the [simulation state](#simulation-state). The calculations for each vehicle are done separately. These calculations are divided in different [stages](#stages). The [control loop](#control-loop) makes sure that all the calculations are consistent by creating __synchronization barriers__ in between stages. No one moves to the following stage before the calculations for all the vehicles are finished in the current one. Each vehicle has to go through the following stages:
TM vehicles do not have a predefined route, and path choices are taken randomly at junctions. Having this in mind, the [In-Memory Map](#in-memory-map) simplifies the map as a grid of waypoints, and a near-future path to follow is created as a list of waypoints ahead. The path of every vehicle will be stored by the [PBVT](#pbvt) component (Path Buffers & Vehicle Tracking), so that these can be accessed easily and modified in future stages.
During this stage, bounding boxes are extended over the path of each vehicle to identify potential collision hazards, which are then managed when necessary.
Similar to the Collision Stage, this stage identifies potential hazards that affect the path of the vehicle according to traffic light influence, stop signs, and junction priority.
Once a path has been defined, this stage computes vehicle movement. A [PID controller](#pid-controller) is used to determine how to reach the target values. This movement is then translated into an actual CARLA command to be applied.
Finally, the TM has calculated the next command for every vehicle, and now it is only a matter of applying them. All the commands are gathered by the [command array](#command-array), and sent to the CARLA server in a batch so that they are applied in the same frame.
And thus the cycle is concluded. The TM follows this logic on every step of the simulation. For a better understanding of the role that each component plays, this section includes an in-depth description of each of them.
* Scans the world to keep track of all the vehicles and walkers in it, their positions and velocities. If physics are enabled, the velocity is retrieved by [Vehicle.get_velocity()](python_api.md#carla.Vehicle). Instead, if physics are disabled, the velocity is computed using the history of position updates over time.
* Stores the position, velocity and additional information (traffic light influence, bounding boxes, etc) of every vehicle and walker in the [simulation state](#simulation-state) module.
* Sends the batch to the CARLA server. Either __apply_batch()__ or __apply_batch_synch()__ in [carla.Client](../python_api/#carla.Client) will be called, depending if the simulation is running in asynchronous or synchronous mode, respectively.
* Synchronization barriers are placed between stages so that consistency is guaranteed. Calculations for all vehicles must finish before any of them moves to the next stage, ensuring that all vehicles are updated in the same frame.
* Coordinates the transition between [stages](#stages) so that all the calculations are done in sync.
* When the last stage ([Motion Planner Stage](#stage-4-motion-planner-stage)) finishes, the [command array](#command-array) is sent to the server. In this way, there are no frame delays between the calculations of the control loop, and the commands being applied.
PBVT stands for __Path Buffer and Vehicle Tracking__. This data structure contains the expected path for every vehicle so that it can be easily accessible during the [control loop](#control-loop).
* Contains the [In-Memory Map](#in-memory-map) that will be used by the [Localization Stage](#stage-1-localization-stage) to relate every vehicle to the nearest waypoint and possible overlapping paths.
* Using the information gathered by the [Motion Planner Stage](#stage-4-motion-planner-stage), estimates the throttle, brake and steering input needed to reach a target value.
* The adjustment is made depending on the specific parameterization of the controller, which can be modified if desired. Read more about [PID controllers](https://en.wikipedia.org/wiki/PID_controller) to learn how to do it.
* Receives the current state of all vehicles and walkers in the world from the [ALSM](#alsm), including their position, velocity and some additional information (such as traffic light influence and state). It also stores some additional information such as whether these vehicles are under the influence of a traffic light and what the current state of that traffic light is.
* Obtains the position and velocity of all the vehicles from the [simulation state](#simulation-state).
* Using the [In-Memory Map](#in-memory-map), relates every vehicle with a list of waypoints that describes its current location and near-future path, according to its trajectory. The faster the vehicle goes, the larger the list will be.
* For every pair, extends bounding boxes along the path ahead (geodesic boundaries), to check if they actually overlap and the risk of collision is real.
Third stage in the [control loop](#control-loop). Triggers hazards to follow traffic regulations such as traffic lights, stop signs, and priority at junctions.
* If the vehicle is under the influence of a yellow or red traffic light or a stop sign, sets a traffic hazard.
* If the vehicle is in a non-signalized junction, a bounding box is extended along its path. Vehicles with overlapping paths follow a FIFO order to move and waits are set to a fixed time.
* Gathers all the information so far: position and velocity of the vehicles ([simulation state](#simulation-state)), their path ([PBVT](#pbvt)) and hazards ([Collision Stage](#stage-2-collision-stage) and [Traffic Light Stage](#stage-3-traffic-light-stage)).
* Makes high-level decisions about how the vehicle should move, for example, computing the brake needed to prevent a collision hazard. A [PID controller](#pid-controller) is used to estimate behaviors according to target values.
First of all there are some general behaviour patterns the TM will generate that should be understood beforehand. These statements are inherent to the way the TM is implemented:
* __Vehicles are not goal-oriented,__ they follow a trajectory and whenever approaching a junction, choose a path randomly. Their path is endless and they will never stop roaming around the city.
* __Vehicles' target speed is 70% their current speed limit,__ unless any other value is set.
* __Junction priority does not follow traffic regulations.__ The TM uses its own priority system while junction complexity is solved. This may cause some issues such as a vehicle inside a roundabout yielding to a vehicle trying to get in.
The TM provides a set of possibilities so the user can establish specific behaviours. All the methods accessible from the Python API are listed in the [documentation](../python_api/#carla.TrafficManager). Here is a general summary of what the current possibilities are.
| **Safety conditions:** | <br>**1\.** Set a minimum distance between stopped vehicles (for a single vehicle or for all of them). This will affect the minimum moving distance. <br>**2\.** Set an intended speed regarding current speed limitation (for a single vehicle or for all of them). <br>**3\.** Reset traffic lights. |
| **Collision managing:** | <br>**1\.** Enable/Disable collisions between a vehicle and a specific actor. <br>**2\.** Make a vehicle ignore all other vehicles. <br>**3\.** Make a vehicle ignore all walkers. <br>**4\.** Make a vehicle ignore all traffic lights. |
Now the TM needs some vehicles to be in charge of. In order to do so, enable the autopilot mode for the set of vehicles to be managed. Retrieve the port of the TM object that has been created. If no port is provided, it will try to connect to a TM in the default port, `8000`. If the TM does not exist, it will create it.
In multiclient situations, creating or connecting to a TM is not that straightforward. Take a look at the [*Running multiple Traffic Managers*](#running-multiple-traffic-managers) section to learn more about this.
The script `spawn_npc.py` in `/PythonAPI/examples` creates a TM instance in the port passed as an argument and registers every vehicle spawned to it by setting the autopilot to True on a batch.
The following example creates an instance of the TM and sets a dangerous behaviour for a specific car that will ignore all traffic lights, leave no safety distance with the rest and drive at 120% its current speed limit.
Now, here is an example that registers that same list of vehicles but instead is set to conduct them with a moderate behaviour. The vehicles will drive at 80% their current speed limit, leaving at least 5 meters between them and never perform a lane change.
The TM is not an actor that needs to be destroyed, it will stop when the corresponding client stops. This is automatically managed by the API so the user does not have to take care of it.
However, it is important that when shutting down a TM, the vehicles registered to it are destroyed. Otherwise, they will stop in place, as no one will be conducting them. The script `spawn_npc.py` does this automatically.
Shutting down a __TM-Server__ will shut down the __TM-Clients__ connecting to it. To learn the difference between a __TM-Server__ and a __TM-Client__ read about [*Running multiple Traffic Managers*](#running-multiple-traffic-managers).
In deterministic mode, the Traffic Manager will always produce the same results and behaviours under the same conditions. Do not mistake determinism with the recorder. While the recorder allows you to store the log of a simulation to play it back, determinism ensures that Traffic Manager will always have the same output over different executions of a script, if the same conditions are maintained.
Deterministic mode is meant to be used __in synchronous mode only__. In asynchronous mode, there is much less control over the simulation, and determinism cannot be achieved. Read the considerations to run [TM in synchronous mode](#synchronous-mode) before using it.
To enable deterministic mode, simply call the following method in your script.
```py
my_tm.set_random_device_seed(seed_value)
```
`seed_value` is an `int` number from which all the random numbers will be generated. The value is not relevant itself, but the same value will always result in the same output. Two simulations, with the same conditions, that use the same seed value, will be deterministic.
The deterministic mode can be tested when using the `spawn_npc.py` example script, using a simple argument. The following example sets the seed to `9` for no specific reason.
```sh
cd PythonAPI/examples
python3 spawn_npc.py -n 50 --sync --seed 9
```
!!! Warning
Make sure to set both the world and the TM to synchronous mode before enabling deterministic mode.
In hybrid mode, either all vehicle physics can be disabled, or enabled only in a radius around an ego vehicle with the tag `hero`. This feature removes the vehicle physics bottleneck from the simulator. Since vehicle physics are not active, all cars move by teleportation. This feature relies on [Actor.set_simulate_physics()](https://carla.readthedocs.io/en/latest/python_api/#carla.Actor.set_simulate_physics). However, not all the physics are disregarded. Basic calculations for a linear acceleration are maintained. By doing so, the position update, and vehicle speed still look realistic. This guarantees that when a vehicle enables or disables its physics, the transition is fluid.
* [__TrafficManager.set_hybrid_physics_mode(True)__](https://carla.readthedocs.io/en/latest/python_api/#carla.TrafficManager.set_hybrid_physics_mode) — This method will enable the hybrid mode for the Traffic Manager object calling it.
*__Running `spawn_npc.py` with the flag `--hybrid`__ — The vehicles spawned will be registered to a Traffic Manager stated inside the script, and this will run with the hybrid physics on.
*__Radius__*(default = 50 meters)* — States the proximity area around the ego vehicle where physics are enabled. The value can be changed via [traffic_manager.set_hybrid_physics_radius(r)](python_api.md#carla.TrafficManager.set_hybrid_physics_radius).
*__Ego vehicle__ — A vehicle tagged with `role_name='hero'` that will act as the center of the radius.
*__If there is none,__ all the vehicles will disable physics.
*__If there are many,__ the radius will be considered for all of them. That will create different areas of influence with physics enabled.
The following example shows how the physics are enabled and disabled when hybrid mode is on. The __ego vehicle__ is tagged with a __red square__. Vehicles with __physics disabled__ are tagged with a __blue square__. When inside the area of influence stated by the radius, __physics are enabled and the tag becomes green__.
When working with different clients containing different TMs, understanding the inner implementation of the TM in the client-server architecture becomes especially relevant. There is one key to ruling these scenarios: __the port__.
A client creates a TM by communicating with the server and passing the intended port to be used for that purpose. The port can either be stated or not, using the default of `8000`.
* __TM-Server__ — The port is free. This type of TM is in charge of its own logic, managed in `TrafficManagerLocal.cpp`. The following code creates two TM-Servers. Each one connects to a different port, not previously used.
* __TM-Client__ — The port is occupied by another TM. These instances are not in charge of their own logic. Instead, they ask for changes in the parameters of the __TM-Server__ they are connected to in `TrafficManagerRemote.cpp`. The following code creates two TM-Clients, that connect with the TM-Servers previously created.
Note how the default creation of a TM uses always `port=8000`, and so, only the first time a __TM-Server__ is created. The rest will be __TM-Clients__ connecting to it.
The CARLA server keeps a register of all the TM instances internally by storing the port and the client IP (hidden to the user) that link to them. Right now there is no way to check the TM instances that have been created so far. A connection will always be attempted when trying to create an instance and it will either create a new __TM-Server__ or a __TM-Client__.
Multisimulation is when there are more than one CARLA server running at the same time. The TM declaration is not relevant. As long as the computational power allows for it, the TM can run multiple simulations at a time without any problems.
The concept of multisimulation is independent from the Traffic Manager itself. The example above runs two CARLA simulations in parallel, A and B. In each of them, a TM-Server is created independently from the other. Simulation A could run a Multiclient TM while simulation B is running a MultiTM, or no TM at all.
The only possible issue arising from this is a client trying to connect to an already existing TM which is not running on the selected simulation. In case this happens, an error message will appear and the connection will be aborted to prevent interferences between simulations.
The TM is a module constantly evolving and trying to adapt the range of possibilities that it presents. For instance, in order to get more realistic behaviours we can have many clients with different TM in charge of sets of vehicles with specific and distinct behaviours. This range of possibilities also makes for a lot of different configurations that can get really complex and specific. For this reason, the following are considerations that should be taken into account when working with the TM:
If the CARLA server is set to synchronous mode, the Traffic Manager must be set to synchronous mode too. To do so, your script should be similar to the following:
* In a __[multiclient](#multiclient)__ situation, only the __TM-Server__ must be set to synchronous mode.
* In a __[multiTM](#multitm)__ situation, only __one of the TM-Server__ must be set to synchronous mode.
* The __[ScenarioRunner module](https://carla-scenariorunner.readthedocs.io/en/latest/)__, already runs a TM. In this case, the TM inside ScenarioRunner will be the one set to sync mode.
Disable the synchronous mode (both, world and TM sync mode) in the script doing the ticks before it finishes. Otherwise, the server will be blocked, waiting forever for a tick.
The Traffic Manager is one of the most complex features in CARLA and so, one that is prone to all kind of unexpected and really specific issues. The CARLA forum is open to everybody to post any doubts or suggestions, and it is the best way to keep track of issues and help the CARLA community to become greater. Feel free to login and join the community.
