The design for the board to supply power to both the Pi Zero and robot modules is fairly simple. It’s two components are a single-cell, LiPo charging circuit based on the MCP73831 and the Pololu adjustable step-up regulator. The charging circuit is fairly straightforward so I’ll just explain the step-up circuit which takes the output voltage from the single cell battery, passes it to the four pin input side of the step-up regulator, and gets back a boosted voltage from the second four pin output side of the regulator.
I also added four beefy diodes between the regulator and the 5V supply for the system. My reasoning behind this was to provide the ability to use multiple power boards, and therefore multiple LiPo batteries, in the system. Each regulator is limited to roughly three amps of current with the multiple boards, Arduinos, and the Pi Zero taking up a fair chunk of that. I was worried that the regulator wouldn’t be able to supply enough current to the motors and so I added the diodes so that multiple power boards could be safely put in parallel. I also decided to use an adjustable voltage regulator so that the output voltage of the regulator could be set higher than 5V so that the actual voltage seen after the diodes would be 5V after the diode forward voltage.
The first board I’ll go over is the Motor Board prototype. It’s fairly simple; excluding the Arduino that handles communication and control it only consists of encoder connectors and a dual H-bridge. My planned use for this prototype will be to control the main drive motors of the robot base module. The schematic and PCB are shown below.
I’m using an L293D chip for the motor driver. As shown in Figure 10 of the datasheet, each half of the chip can function as a bidirectional motor controller. By driving PWM signals to the control inputs of the chip, the speed of the motors can also be controlled. Unfortunately, each half of the bridge can only handle up to 600mA, which is relatively low, but it’ll sufficient for controlling the basic motors I plan on using on the prototype robot.
The encoders I’m using are KY-040 Rotary Encoders which will measure the number of rotations of the motors and provide feedback to the motor controller. The linked description explains how the encoders work better than I can, but essentially the Arduino will measure the speed and number of rotations that the motor is travelling at and apply more or less current to the motor to get it to the desired end position. I plan on using a basic PID loop for this which I will cover in a later post.
In addition to the encoder connectors on the left, I’ve also added a basic debounce circuit in the top left of the schematic. Because these encoders use mechanical switches, they’re subject to mechanical bouncing of the switch contacts and so I plan on using a capacitor to act as a low pass filter and absorb these bounces and clean up the encoder signal.
Before doing the circuit design and PCB layout I wanted to briefly outline the architecture of the core robot modules and the communication between them. The core modules are the ones that will be present on every robot in the swarm and provides the critical functionality that’s required for the robot to operate. To make programming easier and reduce load on the main processor I’ve decided that each module will have a processor that will intelligently communicate with the main board and abstract away as many unnecessary details as possible.
The main board will be a single board computer that will do all of the “thinking” for the whole robot. It will handle communication to the swarm master and pass information between each of the modules within a robot. Due to the large community and low cost I’ve decided to use the Pi Zero W as the main processor. I’ll be using the GPIO header as the main connector and each module will be connected together through this common header. Each of the submodules will be I2C slaves and will be directed by the Raspberry Pi on the I2C bus. The Pi Zero W also has the benefit of built-in WiFi which further reduces the cost and complexity of the robot.
The power module will be fairly dumb with it’s only task being to supply power to the whole module stack. For the first iteration I only plan on measuring battery voltage/percentage via the onboard processor but I may expand these capabilities in the future to include things like power usage, current, battery health, etc.
The motor controller board will be responsible for the actual movement of the robot. My initial design will have the ability to control two motors as well as encoders to close the loop and verify that the robot has moved where expected. As I mentioned in my previous post, encoders are subject to drift so these encoders will verify that the robot movement is in the right ball park and will only be used for a single move. This means that the encoder position will be reset after each move is completed.
I plan on having this module be intelligent enough to handle all coordinated motion without intervention from the main processor board. The Raspberry Pi should be able to send the board an XY position, say “Go here,” and have the motor board handle the rest. It will also be able to query this submodule for certain status such as current position, status of the move, etc.
Due to the relative complexity of the localization algorithm, as outlined in my previous post, the localization board will be the most complicated. It required IR LEDs for transmitting the sync signal, IR receivers for receiving another robot’s sync signal, and ultrasonic transducers for sending and receiving localization pings.
This submodule won’t be as independent as the motor controller board due to the inter-robot communication requirements of the localization algorithm, but the interface should be simple. The board will only take one command, to send out a ping, so it knows to trigger the sync signal and localization signal on the IR LED and ultrasonic transmitter. Otherwise the board will constantly wait to receive an IR sync signal and will calculate distance from the transmitting robot based on how long it is until the localization signal arrives. The main processor board will then be able to query the localization board for the distance to the transmitting robot.
With the basics of the submodules laid out the next step is designing the circuits and PCBs!
One of the biggest problems with precise motion in robotics is localization. Encoders are only so accurate and tend to drift over time. Absolute positioning systems like GPS have error margins measured in meters. The solutions to this is often complicated (e.g. SLAM) or expensive (e.g. RTK GPS). As a cheap and simple alternative to this, I’m going to attempt to use incremental trilateration to enable the robots to determine their own position relative to their siblings.
Why It’s Important
An accurate map of where the robots are is critical for proper planning of their movements. Without it they could get in each other’s way or fall off of a cliff! It’s also necessary for them to perform tasks that require precise motion such as moving debris or laying brick. While the robots will also be teleoperated and have an onboard camera, it’s hard for the human eye to properly determine depth and estimate measurements from a 2D image.
The most common method for robot positioning is using encoders to measure the distance the wheels have traveled by counting rotations. There are some caveats to this method, however. If a wheel slips the encoder’s accuracy will suffer and it will think it’s traveled farther than it has. Over time the encoder’s measurement will drift farther and farther from reality. This problem will only be exacerbated by the rough and slippery terrain that these robots will eventually be operating on. However, with incremental trilateration the robot will recalculate an absolute position every time it moves. This absolute measurement won’t be susceptible to location drift.
The method I’m proposing for robot localization is based on trilateration, which is the same method GPS satellites use to determine position. This method, however, will be able to attain greater accuracy with cheaper and less complex hardware. Rather than the speed-of-light radio signals that GPS satellites use for trilateration, I plan on using much slower sound waves to do the same calculations. This makes it possible to do signal detection and calculation with a relatively slow microcontroller instead of high speed DSPs and custom silicon. I also plan on having an IR blast in the beginning to synchronize all of the robots before each sound pulse is sent. This sync signal will provide a trigger to the robots acting as base points so that they don’t always need to be waiting for sound pulses.
Incremental trilateration consists of four steps:
For this method to work the robots need to start off at set points. This means that the robot master and two of the robots need to start at known locations. Trilateration requires the knowledge of three base points for calculating the fourth point and so absolute positions of the three base points need to be known.
The animation above covers the movement, sync transmit, and signal transmit steps of the process. The big square and two smaller circles on the sides represent the base station and two of the robots at known points.
In this step the robot that’s currently active moves, either towards a target or to explore. In the example above the robot in front moves along the Y axis from its position directly in front of the base station to an indeterminate position in front and to the right of where it was before.
The expanding, red circle sent out from the robot represents a sync signal that will be sent from the robot at an unknown position to the three other robots at known locations. In my planned implementation this will be a 40Khz IR signal. Once the three receiving robots receive this signal they know to start counting and waiting for the signal transmit. It should be noted that I’m ignoring the travel time of the IR pulse because the speed of light is so fast that it can be considered insignificant over the small distances that the robots will be moving.
Once the sync IR signal has been received, the robots will start counting until they see the sound pulse. I plan on using 40Khz ultrasonic transducers to handle this and generate an inaudible sound wave. Once the robots see the sound pulse they will stop counting and save the difference in time between the sync and signal transmissions. Using the speed of sound they can then calculate the distance to the robot that sent the transmissions.
Step 4: Calculation
Once each of the base robots calculates the distance to the moving robot they can then effectively draw a circle around themselves with a radius of the distance they’ve calculated. The base robots can definitively know that the moving robot is somewhere on the edge of the circle.
Using the circles drawn by each of the base robots and the known direction that the robot traveled in, it can then be determined that the moving robot is sitting at the point where each of the three circles overlap, as displayed in the image above.
There are a few limitations to the incremental trilateration method that I’d like to explain and propose solutions to.
The first and most inconvenient is that the robots will require known initial conditions. This means that at least three nodes in the localization network need to be at predefined positions, otherwise it’s impossible to calculate the position of the moving robot. This makes setup a little harder and introduces accuracy problems if the initial conditions aren’t perfect.
Some of this can be mitigated by the fact that each of the nodes can determine the location to the master node using the sync and signal method. If each robot is placed along a single straight line (which can be considered a line perpendicular to the Y axis at a known value of Y), it can send sync and signal transmissions to the master to determine its X offset.
Another possibility would be adding three IR and ultrasonic receivers to the master at predefined locations so that the master itself can act as the three reference points for the moving robot. This introduces some complexity but may ultimately be worth it.
Turn based movement
Another limitation is that in the above scenario with four nodes, only one robot can move at a time as it needs the three reference points to be stationary. This is less limiting in larger networks as the necessity of three reference points means (N – 3) robots can be moving with N being the number of robots in the network. For large values of N the limitation is less. However, because the IR and ultrasound are using the air as a common bus the actual transmissions will need to be kept to one at a time to prevent collisions.
Accuracy is the biggest concern with this system. Until I test this I won’t be sure the exact accuracy of the system but there is a lot of variability that can cause problems. The reason I’m using ultrasound for the signal transmission instead of light is because the speed of light is much too fast and the clock speed of microprocessors is much too low to properly detect the signal over small distances. However the ultrasound still has the same limitations, albeit to a lesser degree.
The Arduino micros timer measurement has a resolution of four microseconds. Since the speed of sound is 343 m/s, the ultrasonic pulse will be able to travel 0.001372 meters (343 m/s * 4e-6 seconds) or 1.372 millimeters for one increment of the micros() counter. This is only the maximum theoretical resolution, however, since it doesn’t take into account things such as digital read or sensor latency. Ultimately the actual resolution is something I’ll have to determine experimentally. I’m hoping for a 1cm accuracy for my initial implementation and will have to search for optimizations should that not be immediately achievable.
Another thing to take into account is that the speed of sound changes with temperature. However this can be fixed by using a temperature sensor to more accurately calculate the speed of sound as shown here.
Next up is planning out and designing the actual hardware!
Before diving into the nitty-gritty of the robot design I wanted to take a moment and lay out a brief description of the network and base robot architecture. Below is a block diagram and description of both the network topology and the configuration of the robots.
The pure swarm approach for a fully modular network would involve a mesh network with with no centralized control source. Instead of doing this I’m opting for a robot swarm with central control node, much like how a bee hive has a queen. And, since robot dancing has not reached bee levels of communication, I’m going to leverage existing technology and use WiFi. This will make it easier to use existing single board computers (e.g. the Raspberry Pi) for robot control instead of making a homebrew control board. Having all of the robots on a single WiFi network will also make it easy to remote login to the individual robots for telepresence control.
I also plan on having a single, central node to handle the complex control and collate all of the data provided by the individual swarm robots. This central point will also have the WiFi access point. Having a single master may seem counter-intuitive for a robot swarm but it will make development and control much easier. A powerful central computer can do complicated operations such as image processing and parsing the most efficient paths for each robot to take. The central node can also handle delegation by directing each robot to assume a role in heterogeneous swarms when different tasks need to be handled by the robots. Another benefit of making a single master node is debugging. Having the central node keep track of all of the data will make it easier to access the swarm’s status and provide a clear picture of how the system is operating.
Base Robot Configuration
I’ve tried my best to limit the robot’s base design (i.e., the components that will be common between all robots no matter their role or attached modules) to the very minimum. The block diagram above is what I came up with.
Each robot will have three core modules built in to the base design: one for power control and distribution, one for motor control, and one for localization. The only exception will be the central node which won’t have the motor control module. The modules will all report to a central CPU. Each of these modules will be intelligent with its own microcontroller for real time control and calculations. Having a microcontroller for each module abstracts away the processing required for each module and lets the CPU retrieve processed data and send commands without having to manage every single component.
I decided on using a full SoC rather than just a microcontroller to help speed up development and to potentially allow for some image processing and other calculations to be done locally. The Pi Zero W seems like the best bet for this at the moment due to its native camera support and on-board WiFi (and the large support community is a huge plus!). Using a full Linux system will also make software design easier without requiring constant retrieval and firmware flashing every time the software changes. It will be simple enough to remote in to each robot for control and status update over SSH.
I plan on defining the interface and functionality of each module in the next few posts. I also want to outline and explain my method for robot localization in detail. That will be my next post.
Infrastructure and construction robots are a group of modular robotic agents that can work cooperatively to automate various tasks such as construction, inspection, and repair. Each robot is capable of multiple functions, as capacities can be modified through the installation of different attachments on the unit. Each base unit is designed to be low cost to minimize the time and monetary investment and protect against significant losses due to the failure of a single robot. A group of robots can be teleoperated to increase the efficiency of skilled operators or can be given basic directives to automate simple or repetitive tasks.
Problem A: Infrastructure
U.S. infrastructure is currently rated as subpar and is continuously degrading (Graded D+ by ASCE)
Part of the issue lies in the extreme costs of repairs necessary across the nation (estimated $4.59 billion by 2025 to fix)
This massive increase in spending would cause a drag on the economy ($3.9 trillion GDP loss by 2025)
Too many problems, too expensive, takes too long
Problem B: Disaster Relief Management
Disasters are difficult to recover from and each disaster has unique requirements and problems that need to be addressed
Disasters are inherently unpredictable and so disaster preparation would require addressing all possible problems before it occurs, a prospect that can be prohibitively expensive
It’s also expensive to be reactionary and ship required relief materials for every disaster only after the disaster has occurred
Loss of life is worse immediately following the disaster before the relief supplies can make it to their destination
Massive, repeated shipments can be problematic due to bureaucratic and logistical delays
Problem C: Construction and Landscaping
Building homes and structures is expensive and time consuming
Machines for construction are expensive and have high specialization sometimes requiring many different machines for a single project
Cheaper, mass produced materials result in low levels of customization and customer satisfaction also requiring expensive shipment to the destination
Heavy earth moving machinery is often needed for landscaping projects which can be expensive or difficult to acquire
Using heavy machinery can be difficult and requires additional skills; alternatively hiring professionals only makes the work more expensive
Problem D: Non-Terrestrial Construction
Sending up satellites is limiting in both space and weight
After the heavy costs associated with initial construction, there are currently no options for repairs in the case of damage
Space junk is a problem without a good method of deorbiting debris
Expensive and restrictive to need to send all materials to construct a habitat on another planet
Materials to build habitats exist on planets but no machines currently capable of utilizing them or constructing habitats
Remote operated requiring skilled workers or semi-autonomous task completion
Perform checking and eventually construction that’s more cost effective
Makes workers more efficient and amplifies their capabilities
Increase in safety with cheap, expendable robots performing dangerous work
Flexible function using modular attachments means each robot is capable of many different actions
Small size means robots can work concurrently in the same location, reducing multiple phases of a project into a single, incremental step
Swarm methodology for a group of robots means tasks can be accomplished much quicker
A combination of a large swarm and the modularity means specific tasks can be dynamically allocated due to shifting requirements
Robot with modular attachments that enable it to serve various functions
Cheap base model that can be produced in quantity and has basic motion, communication, and sensor capabilities
Assignable roles so that broken robots can easily be replaced by spare units or units can be re-assigned based on need
Teleoperation with visual data so that users can get direct feedback about the status of the project’s target
Limited physical intervention required by users
Ability to direct movement and basics tasks to be performed by the robots
Modular attachments can be swapped autonomously
Dynamic role allocation so that the robots automatically determine the best distribution to get a task done efficiently
Coordinated motion so the robots can be given simple direction to accomplish group behavior
Machine learning can automatically flag problematic inspection data
Augmented reality data so that the users can see the project target and the current progress side by side
Advanced autonomy with the robots having the capabilities to perform a great deal of tasks with little user control required
Camera for teleoperation and visual data collection
Central communication node to route mesh network packets
Master node compiles debug information and swarm state
Master establishes swarm requirements and passes them on to individual agents
Handled over WiFi with the master node providing the central access point
Communication through a web interface hosted on the central node
Accessible by connecting to the swarm network
High level control abilities to direct general motion of the group and assign tasks
Lower level control also available for finer motion control
Individual robots are selectable so that debug information, robot state, and individual command interfaces are available
Jobs are dynamically allocated and passed on to the individual robots via communication with the master node
Robots automatically connect to the attachments required to perform their assigned tasks
Gripper: Movement of structural material
Dumper: Transportation and removal of debris, earth, sand, etc.
Arm: Fine manipulation of objects
1. First Functional Prototype
Single robotic unit
Direct control basic interface (command line, GPS coordinates, etc.)
Basic motion and movement commands with low accuracy
2. Attachment Prototypes
Several basic, modular attachments for the prototype robot so it can perform various functions (e.g. gripper, arm, digger, screwdriver, drill, material transportation)
Automatic identification and control of the different attachments
3. Prototype Swarm
Multiple units controllable via a single interface
Different attachments to show multi-use cooperation
Direct control basic interface for individual units as well as group movements
Central communication hub to route communication between robots and provide single access point that distributes commands to individual swarm members
4. Survey and Analysis Demo
Demonstration of basic survey capabilities using direct control interface and teleoperated swarm
Robots coordinate with each other to map out a structure and provide detailed pictures
Optional additional sensors for measuring other useful data (e.g. radiation, temperature, vibration)
5. Repair Demo
Demonstration of basic repair capabilities using direct control interface and teleoperated swarm
Robots are capable of moving repair materials into place and performing the repairs without physical operator intervention
Robots can cooperate as a cohesive group to transfer repair materials and/or remove broken material
6. Construction Demo
Demonstration of basic construction capabilities using direct control interface and teleoperated swarm
Robots are capable of assembling an entirely new structure without physical operator intervention
Robots are capable of working together to prepare construction area and build a structure
New project alert! I was throwing around some multidisciplinary project ideas with two Mechanical Engineering friends and we talked about having some sort of robotic swarm for construction, disaster relief, infrastructure inspection, etc. Right now it’s not really a fully formed idea but I did submit it to the Hackaday Prize contest anyway since they’re currently running an “idea” seed funding phase. Here’s the project link. I think this idea has a lot of potential and I’m hoping this can grow into something that’s useful and beneficial to society. More details to follow!
It’s been a while since I’ve posted here, not because I haven’t been working on projects but more because I’ve spent the better part of a year trying to add some diversity to my hobbies. Last April I decided to finally lose the weight I’d gained since high school so I did a little self-improvement and lost forty pounds while exploring some other interests. Aside from the weight loss I took up running and biking, started learning to play violin, wrote a novel for NaNoWriMo, and am currently experimenting with gardening and cooking. Luckily many of these activities provide opportunities for electronics hacks and designs so I anticipate a few smaller hacks centered around these new hobbies in addition to the large, ongoing projects I have going.
Since I last posted, I’ve redesigned the OpenSCAD to make the robot much smaller. I haven’t documented any of this and I’m pretty sure it’s a dead end, but I’ll try and write up the changes I’ve made and my results soon. I haven’t really done anything else on the PiGRRL Switch. Everything I need is sitting on my desk, I just need to buckle down and put it all together. I plan on starting with the console and verifying it works with a regular USB controller and then finishing the custom controllers.
So I didn’t get a chance to take a video of the controller in action, but I do have a picture of me playing the original Zelda using my soldered up controller prototype. The Raspberry Pi and LCD is hooked up to one battery bank and the controller is hooked up to a separate one. Going through this test I noticed a few things that can be improved.
For starters, I had a hellish time getting enough power to both the Raspberry Pi and LCD. For whatever reason the power supply didn’t seem to be able to give enough juice to keep the Raspberry Pi from rebooting. The display has two micro USB ports on it that are used to power the screen and provide touchscreen feedback to the Pi. However since I currently don’t need to use the touchscreen I just powered the screen directly from the battery bank. To circumvent these power issues in the future I’ll need to see if I can get the Pi to provide more current from the USB ports. If that doesn’t work I can use one of the micro USB ports for touchscreen feedback and cut the red power wire to prevent the screen from drawing power from that port and use the second USB port to power the screen directly from the battery bank.
Another issue I noticed with my controller is the lack of menu buttons. The RetroPie interface requires using Start and Select buttons for navigation, so I had to map those to an attached keyboard since I only have four action buttons and four directional buttons on my prototype. Additionally I’ll also want shoulder buttons since the later consoles all have at least two and the PS1 has four.
The next step for the controller interface is designing the PCB!
The next step in getting the controller to work as a RetroPie controller is writing a driver so the console interprets the serial data coming in from the Bluetooth module the same way it would interpret a USB controller.
I got this idea from petRockBlog who mapped their custom, two-player joystick to two separate virtual controllers in RetroPie. Their code is available on GitHub. After a quick review, I saw that their code is using a C user input emulation library called uinput. In the interest of simplicity I found a Python uinput library and decided to use that to emulate my controller driver.
Below is a full snippet of the code from the GitHub repository followed by an explanation of how it works:
This is the definition of the Controller class. A static variable called NumButtons is used to show the maximum number of buttons a Controller has. A Controller instance is initialized with a path to the Bluetooth serial device which it then connects to. The initialization also creates a uinput Device that it will use to simulate a virtual controller for providing input to the console.
An array of keys is created to represent the possible input keys that will be passed into the system. Right now I’m just using the letters of the alphabet. RetroPie offers full controller customization so the actual keys used don’t really matter. I still plan on changing the controller driver to emulate a joystick with analog controls, however, so this method will need to be updated in the future.
This array will store the device names for all of the Bluetooth serial consoles. At some point I’d like to autodetect all of the serial consoles from the controllers but for now a hardcoded array will work.
The program creates a controller for each serial console in the ControllerNames array and stores them in the Controllers array. It also resets the serial input buffer so old input data doesn’t bog down the driver and it can start processing the most current data.
My driver operates on button transitions (unpressed->pressed and pressed->unpressed) rather than operating on the button state. This is to be consistent with keyboards which work in the same way. Because of this transition-based processing, the previous input packet needs to be stored so the previous packet can be compared to the current packet and differences between the two can be detected. Here the previous packet is instantiated as a dictionary and is initialized to a known, totally unpressed state.
This is the code for the key ‘K’ which is the only type of data I currently have. It iterates over each character in the eight bytes of data and compares it against the data from the previous packet. If the value went to ‘1’ from ‘0’ it emits an event to signal that the key was pressed and if the value went to ‘0’ from ‘1’ it emits an event to signal the opposite.