Swarm Robot Project Simulator
The simulator is a part of the Swarm Robot Project that attempts to model the robots using MATLAB. It is built around a modular control structure so that it can be easily adapted to run other algorithms. Features include asynchronous update cycles, a packet model, lossy communication, and flexible rendering options.
There are now two versions of the simulator that implement different algorithms. They both use the same control structure, which this article outlines in context of the formation control algorithm.
Here the robots orient themselves so that their moments (quantifiers of the group's shape and position) match a predefined set of goal moments.
The swarm will start centered around a random point and then converge on the goal moments (shown as the stationary green ellipse). The user may click a location on the grid to restart the simulation with the robots centered around that point.
Here the robots are placed in an environment with up to 3 variables that are functions of location. The robots collaboratively model these variables and relocate themselves to get optimal sensor readings.
The swarm starts clustered at a random location in an environment with randomly chosen parameters. The three environment variables are displayed on top of each other as red, green, and blue. Click on an agent to open a plot of its error (a black field indicates perfect estimate of all colors). The environment function is defined in basefunction.m and it can be changed to anything.
For this application, the agents' states have three extra data fields:
- Y information matrix
- Tau information vector
- sensorR sensor covariance
To run the simulator:
- Download one of the zip files above.
- Extract the Simulator folder.
- Open simulator.m and press Run.
- When prompted, select "Change Directory".
Each M-file is a MATLAB function. The central one is simulator.m. At the top of simulator.m, you will find some constants (denoted by all-caps) that make it easy to change the performance and rendering of the simulation without having to read into the code. See the inline comments for details.
The control structure represents the decentralized code that runs on the individual e-pucks. It is written modularly and without extraneous dependencies so it can be readily adapted to other algorithms.
The whole control structure is called from updateBotComputations. Information about neighbors flows to and from the consensus estimator via the packet model.
The blue boxes from the flowchart to the right represent functions that are defined in their own files. They are the building blocks of the consensus-based swarm model.
The interface between the agent and the environment. For the formation control implementation, this models an e-puck following a velocity vector.
Consolidates information from the environment into a form readable by Consensus Input and the Controller.
A discrete-time PI estimator that uses packet structs for its input and output.
Injects the agent's own information into the estimator-communication cycle. In the formation control implementation, this step causes the estimates to follow the agents instead of converging to the starting location.
Uses the agent's sensory and consensus information to select a set of controls, sent to the plant.
The velocity output from the Controller is put through a collision avoidance algorithm. It is not shown in the flowchart because the algorithm has access to the positions of all the agents (information not known to any single agent).
Note: the present algorithm does not work very well.
The environment is a set of abstractions that encapsulate the control structure, simulating passage of time, motion of the agents, and lossy wireless communication.
The simulation advances in timesteps; equally spaced spans of time. Timesteps are the unit used in specifying the distance between discrete events, such as updating robot calculations or rendering. Their size (relative to other time-based units) is defined by the constant dT.
The main loop runs once per timestep. It updates the plant every timestep and makes calls to time-dependent update functions as needed.
Since all state data in the simulation can be associated with a robot, the only stored variable is a vector of N robots called bots. Each bot is individually represented as a struct, with the following fields:
- bots vector of N robots
- state information the agent knows about itself
- p position
- angle rotational orientation
- packets vector of N received packets (packet i of bot i is that agent's own estimate and always has an age of 0.)
- age timesteps since received
- v PI decision variable
- w PI estimator state
- u velocity chosen by Controller
- lastUpdate timestep of last update
- waitTime timesteps between last update and next update
- state information the agent knows about itself
Consensus information is transferred between the e-pucks using wireless communications. The simulator transfers packets between robots on an asynchronous basis and includes the possibility of packet loss as a function of distance. Each robot's packets vector contains the most recently received packet from each of its neighbors. A robot will transmit its own consensus information just before its control structure is executed.
Packet loss is modeled by a cumulative Rice function, as advised by The problem of Wi-Fi radio-fading simulation: Solution and applications. As distance between agents increases, the probability of receiving a packet decreases.
Visualizing packet loss
Packet loss can be easily visualized by setting the constant DRAW_CONNECTIONS to true.
Between two agents...
- a black line means at least one agent has a fresh packet from the other.
- a red line means neither has fresh data, but at least one is using stale data.
- no line means neither agent is using data from the other because the last communicated packet is older than PACKET_AGE_LIMIT.
Recording a movie
To record a run of the simulation, set MOVIE_FRAME_COUNT to the desired movie length, in frames. The movie will be saved as "movie.avi" in the same directory as the simulator M-file.
Movies are recorded as uncompressed AVIs because MATLAB does a poor job of compression. See Swarm Robot Project Documentation#Making Videos with Overlays.