No description, website, or topics provided.
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.

This repository contains the realtime kernel, operating system and application used for controlling model trains. It was created over a span of 3 months for the course CS452. Details about the kernel are in

CS 452 Final Project

Name: The Coldwell Kernel

Students: Ryan O'Leary and Elnar Dakeshov

Date: July 27, 2017


The Coldwell Kernel implements preemptive multitasking for the ARM 920T CPU. Tasks may be created, run and exited from. Inter-task communication is based on three synchronization primitives: send, receive and reply. Tasks may await a periodic 10ms timer event or block on UART events. Additionally, a constant time scheduler is implemented with support for up to 32 distinct priorities.

The application supports up to four trains and has been tested with three trains. The trains may be routed to any node on the track without collision. Routing supports reversals to provide the most optimal route. Additionally, trains may enter a free running mode where they drive straight without flipping switches.


        _\/    \/_
          /\__/\          ___ ___      _          __
     _\_\/_/  \_\/_/_    /   /  / /   / \  /   / /_   /   /
      / /\ \__/ /\ \    /__ /__/ /__ /__/ /_/_/ /___ /__ /__
         _/\  /\_       For stickers and glory!
         /\    /\


Git hash: 73e0d74d0953ec920c169f62bb50c890f654d402


In RedBoot, run the following:

> load -b 0x00218000 -h "ARM/rj2olear/coldwell.elf"
> go


$ make <ARGUMENTS>
$ stat build/kernel.elf

Where <ARGUMENTS> may be any combination of the following:

  • CACHE_ENABLED=0: Disables the instruction and data caches
  • OPT_ENABLED=0: Disables optimizations


The following convenient features can be found on the user interface:

  • Clock display in tenths of a second
  • GO/STOP, toggle with Tab key
  • Percent user time spent in the idle task compared to other tasks
  • Top 10 most recent sensor triggers
  • Track A layout including 22 switch states and broken sensors
  • List of reservations for each train
  • Command prompt


Enter instructions at the % prompt. The available instructions are:

  • cal TR SP SEN - stop train on the next trigger.
  • com i [BYTE...] - Send arbitrary bytes over COMi.
  • help - Display this help information.
  • q - Quit and return to RedBoot.
  • route TR SP SEN [OFF] - Route train to sensor at given speed with optional offset.
  • rv NUMBER - Reverse the direction of the train.
  • ssd TR SP NUMBER - Set the stopping distance in mm.
  • sw NUMBER DIR [TR] - Set switch direction ('S' or 'C').
  • task (TID|NAME) - Return info about a task.
  • taskall - Return info about all tasks.
  • tr NUMBER SPEED - Set train speed (0 for stop).

Additionally, the Tab key will stop all trains in case of emergency.



  • multiple trains without collisions
  • reverse trains during routing
  • free running mode with exit reversal
  • set switch with respect to a train

Layers of Abstraction

Each layer is built from multiple classes and/or tasks:

Layer Description
Application There are 3 applications outlined below
Safety Models all trains locations/velocities, performs sensor attribution, performs reservations
Device Handles primitive commands (setSpeed, waitSensors, ...)

For our project, there are three applications:

Application Description
Terminal User types in commands such as tr 24 10
Calibration Used as a tool to find a train's stopping distance at various speeds
Routing Routes a train to the given destination and offset

Why have multiple layers?

  • Software abstraction: New applications can be written without changing the bottom layers.
  • Failure isolation: For example, if the routing algorithm performs a bad route, the safety layer will protect the trains against collision. This allows us to experiment with more sophisticated routing algorithms.

Application Layers


  • Terminal commands are sent to the safety.
  • If a terminal command is unsafe, it will be ignored and a error message is printed.
  • For example: If setting a train's speed will cause a collision, the speed is not increased.


  • Routing finds the shortest path between every node and the end node. This is accomplished using Dijkstra's algorithm where the weight is the length of each track piece.
  • The graph used is the one supplied on the course website. Occasionally, when a sensor or switch is deemed broken, the node's type is modified to relay this information. This way, a broken sensor/switch does not incur any additional cost. Also, the complexity of the graph is not altered.
  • Exposes a function for routing a train from the terminal:
    • First, the routing layer may query the current state of the train. This can be useful for more sophisticated routing.
    • Then, using Dijkstra's algorithm, a shortest path tree is generated in reverse, rooted on the end node.
    • A Gradient Absolute Switch Profile (GASP) is created from the tree. For every switch, the desired direction for passing through the switch is stored. Addtionally, the minimal point (the end switch and offset) is stored.
    • The GASP is passed to the safety server for the given train.
  • After initial routing, there are two reasons for rerouting:
    • Contention: The train has stopped to avoid collision. To resolve this issue, the routing layer generates a new GASP from a graph with the offending node removed.
    • Broken switch: A switch is considered broken if the train diverges from the expected path. To resolve this issue, the routing layer generates a new GASP from a graph with the offending node removed.


  • Calibration is used to find the stopping distance of trains.
  • During calibration, a specific switch is selected. When the train drives over the switch, a stop signal is sent to the train. The stopping distance is measured by hand from the switch's position to the front of the train.
  • Since calibration is in the application layer, a train which is being calibrated will not collide with other trains.

Safety Layer

  • The safety layer exposes the following functions and ensures they are used in a safe manner which will not cause collision:
    • setTrainSpeed(Train, Speed):
    • reverseTrain(Train):
    • setGasp(Train, Gasp): Set all switches with respect to the given train.
    • setSwitch(Train, Switch, SwitchState): Set switch state with respect to the given train.
  • Additionally, a getTrainState function is exposed for getting train state. The state contains:
    • last time a sensor was attributed to the train
    • current speed of the train
    • estimated velocity of the train
    • estimated stopping distance of the train
    • estimated position of the train (graph node and offset)
    • last known exact position of the train (graph node)
    • switch states with respect to the train (GASP)
  • Sensor attribution (see below)
  • Reservations (see below)
  • Train discovery (see below)
  • Models all trains locations/velocities
  • Train state extrapolation

Reservation System (part of the safety layer)

  • Stores reservations: a list of track nodes reserved by each train
    • All trains have reservations regardless of whether they are being routed. This is required to support free-running mode.
    • The reservation starts from the node immediately behind the train to the stopping distance from the next sensor and is contiguous.
    • Whenever the reservation is calculated, the previous reservation is thrown out and recreated with:
      • Starts from the train's location
      • For each next node starting from the trains current node:
        • Reserve each node and its reverse node.
        • If the node is a branch/merge, lookup in the GASP to determine in which direction to route the train.
        • If the node is the first sensor, mark this as the start of the stopping distance.
        • Break once the total distance has exceeded the stopping distance.
      • The backwards reservation functions in the same manner, but the stopping distance is a small constant.
    • If the train is anywhere on a track node within its stopping distance, that track node is considered reserved.
    • No two reservations may contain the same track node
  • Listens to sensors
    • Update train reservations
  • Timeout to prevent contention
    • If a train is blocked on another train's reservation for too long (15s), reverse and signal to routing server to reroute. The routing server is given the offending node.

Attribution System (part of the safety layer)

When a sensor is triggered, it is attributed to the train which has the that sensor in its reservation. Since no two trains may reserve the same sensor, this unambiguously attributes the sensor.

If the sensor is not reserved by any train, it is considered spurious and the "unattributed sensor" error message is printed.

More specifically, the following steps are taken for each sensor notification:

  1. Extrapolate the position of all trains based on time (position & velocity)
  2. Figure out which train caused the sensor (the sensor should be on a train's reservation)
  3. Update that train's position to be exactly over the switch. Also, update the velocity of the train to be more exact.
  4. Inform the reservation class to update reservations.

Train Discovery (part of the safety layer)

When the user sends a command to a train which has never been attributed, a new state is created for the train. The next occurrence of an unattributed sensor is attributed to this new train.

It is impossible to discover two trains at once because there would be ambiguity in the attribution of sensors.


Some of these have meaning in the code, the rest are conceptual.

  • Train Red Zone Hazard: The system has gotten to a state where trains will collide.
  • Train Yellow Zone Hazard: The system has gotten to a state where, if no action is taken, it will enter the red zone imminently. The safety layer will perform an action, such as stopping and/or reversing the train.
  • Switch Red Zone Hazard: A train is over a switch and changing the switch's state may cause the train the derail.
  • Switch Yellow Zone Hazard: A train will soon be over a switch. If a specific direction is desired for the switch, it much be changes imminently before the switch enters the red zone.


This section outlines several challenges and how the provided system solves them:

  • One train following another: The leader behaves normally. The follower is controlled via negative feedback to stay at the cusp of the reserved zone. That is, when a train receives a sensor notification and the offending train's velocity is in the same direction, the follower matches the leader's velocity. To prevent unstable velocities, the following train's speed must always decrease.
  • Two trains on a collision course: There are three generalized situations in which trains may collide:
    • Two trains on a straight: There reserved zones collide and both trains are stopped. After 15s the trains reverse to prevent a deadlock.
    • Two trains on the branches of a switch: One reserved zone is laid down first, the second is blocked. One train will wait for the other.
    • One train on the branch, the other on the merge of a switch: This situation is graphically identical to the straight case.
  • Single failure of sensors: Our reservations are large enough such that they contain the next two sensors. When a sensor fails, the next sensor can be used for attribution.
  • Single failure of a switch: Not solved.
  • Rerouting due to contention
    • Query the state of the train
    • Temporarily remove offending nodes from the graph
    • Creates a path from start to end
    • Tell reservation server to change train's speed

Measuring Velocity

Initially, before two sensors have been attributed for the train, a rough estimate is provided for the velocity.

Therein afterwards, the velocity is calculated as the average of the current estimate and the previous estimate. This means that velocities from previous track segments dimmish exponentially.

Calibrating Stopping Distance

Calibration is used to determine the stopping distance for each train at a given speed. The cal command runs a train at the given speed and sends the stop command at a sensor. The distance is measured between these two reference points:

  • The center of the switch track.
  • The front of the train in the direction the train is driving.

For this reason, we must also measure the stopping distance of each train in reverse.

For accurate results, the average of three trials is taken.

At each sensor, the distance to the end of the route is calculated. With the velocity, it is possible to estimate the stopping time. If the stopping time will pass before the next sensor, a delay is entered to stop the train at the given stopping time.


  • Extra debug information is printed all over the terminal.
  • Sensor modules D and E are printed to the wrong location on the GUI.
  • Trains can only stop properly on Track B.