My little robotic pals

Years ago I decided to build an indoor robot with multiple kinects for navigation and a robotic arm for manipulation. It was an interesting time working out how to do this and what is needed to get a mobile base to map and navigate a static and dynamic indoor space. Any young players reading this might think that ROS can just magically make this all happen. There are some interesting issues to discover building your own base and some, um, "issues" shall we say that you will need to address that are not in the books or docs. I won't spoil it here for the new players other than to say be prepared to be persistent. 

There are two active wheels at the front and a single drag wheel at the back about 12 inches behind the front wheels. I wrote the code to control the arm myself as custom ROS nodes. A great trick here is you can inject sinusoidal movement by injecting a shim ROS node to take one target and smoothly move towards it.

Now I have a new friend for outdoor activity, the "hound bot". The little furry friend is still sans hair but has gps, imu, rc control override, and a ps4 eye camera mounted for depth perception and mapping. Taking a leaf out of one of the big car makers book and only using cameras for navigation. But for me it is about cost since a good lidar is still much to expensive for the hound.

The hound is a sort of monocoque where the copper looking square part at the front is part of a 1/4 inch aircraft grade alloy solid welded chassis that extends the lenght of the robot. The hound can do about 20km/h and is around 20kg in heft. The electronics bay in the middle is protected by a reinforced carbon fibre layup that I did. Mixing material for fun and slight weight loss.

One great part about doing this "because I want to" is that I am unbounded. Academic institutions might say that building robust alloy shells is not a worthwhile task and only the abstract algorithms matter. I get to pick and choose what matters based purely on what is interesting, what is hard to do (yay!), and what will help me get the robot to perform a task that I want.

The hound will get gripper(s) so it can autonomously "fetch" things for me such as the mail or go find and pick up objects on the lawn.

Houndbot suspension test fit

I now have a few crossover plates in the works to hold the upgraded suspension in place. See the front wheel of the robot on your right. The bottom side is held in place with a crossover to go from the beam to a 1/4 inch bearing mount. The high side uses one of the hub mount brackets which are a fairly thick alloy and four pretapped attachment blocks. To that I screw my newly minted alloy blocks which have a sequence of M8 sized holes in them. I was unsure of the final fit on the robot so made three holes to give me vertical variance to help set the suspension in the place that I want.

Notice that the high tensile M8 bolt attached to the top suspension is at a slight angle. In the end the top of the suspension will be between the two new alloy plates. But to do that I need to trim some waste from the plates, but to do that I needed to test mount to see where and what needs to be trimmed. I now have an idea of what to trim for a final test mount ☺.

Below is a close up view of the coil over showing the good clearance from the tire and wheel assembly and the black markings on the top plate giving an idea of the material that I will be removing so that the top tension nut on the suspension clears the plate.

 The mounting hole in the suspension is 8mm diameter. The bearing blocks are for 1/4 inch (~6.35mm) diameters. For test mounting I got some 1/4 inch threaded rod and hacked off about what was needed to get clear of both ends of the assembly. M8 nylock nuts on both sides provide a good first mounting for testing. The crossover plate that I made is secured to the beam by two bolts. At the moment the bearing block is held to the crossover by JB Weld only, I will likely use that to hold the piece and drill through both chunks of ally and bolt them together too. It's somewhat interesting how well these sorts of JB and threaded rod assemblies seem to work though. But a fracture in the adhesive at 20km/h when landing from a jump without a bolt fallback is asking for trouble.

The top mount is shown below. I originally had the shock around the other way, to give maximum clearance at the bottom so the tire didn't touch the shock. But with the bottom mount out this far I flipped the shock to give maximum clearance to the top mounting plates instead.

So now all I need is to cut down the top plates, drill bolt holes for the bearing to crossover plate at the bottom, sand the new bits smooth, and maybe I'll end up using the threaded rod at the bottom with some JB to soak up the difference from 1/4 inch to M8.

Oh, and another order to get the last handful of parts needed for the mounting.

Houndbot rolling stock upgrade

After getting Terry the robot to navigate around inside with multiple Kinects as depth sensors I have now turned my attention to outdoor navigation using two cameras as sensors. The cameras are from a PS4 eye which I hacked to be able to connect to a normal machine. The robot originally used 5.4 inch wheels which were run with foam inside them. This sort of arrangement can be seen in many builds in the Radio Controlled (RC) world and worked well when the robot was simple and fairly light. Now that it is well over 10kg the same RC style build doesn't necessarily still work. Foam compresses a bit to easily.

I have upgraded to 12 inch wheels with air tube tires. This jump seemed a bit risky, would the new setup overwhelm the robot? Once I modified the wheels and came up with an initial mounting scheme to test I think the 12 inch is closer to what the robot naturally wants to have. This should boost the maximum speed of the machine to around 20km/h which is probably as much as you might want on something autonomous. For example, if your robot can out run you things get interesting.

I had to get the wheels attached in order to work out clearances for the suspension upgrade. While the original suspension worked great for a robot that you only add 1-2kg to, with an itx case, two batteries, a fused power supply etc things seem to have added up to too much weight for the springs to counter.

I now have some new small 'coil overs' in hand which are taken from mini mountain bike suspension. They are too heavy for what I am using, with around 600lb/inch compression. I have in mind some places that use coil overs in between the RC ones and the push bike ones which I may end up using. Also with slightly higher travel distance.

As the photo reveals, I don't actually have the new suspension attached yet. I'm thinking about a setup based around two bearing mounts from sparkfun. I'd order from servocity but sfe has cheaper intl shipping :o Anyway, two bearing mounts at the top, two at the bottom and a steel shaft that is 8mm in the middle and 1/4 inch (6.35mm) on the edges. Creating the shafts like that, with the 8mm part just the right length will trap the shaft between the two bearing mounts for me. I might tack weld on either side of the coil over mounts so there is no side to side movement of the suspension.

Yes, hubs and clamping collars were by first thought for the build and would be nice, but a reasonable result for a manageable price is also a factor.

Terry & the start of a video project.

I did a test video showing various parts of Terry the Robot while it was all switched off and talking about each bit as I moved around. Below are some videos of the robot with batteries a humming and a little movement. First up is a fairly dark room and a display of what things look like just using the lighting from the robot itself. All the blinking arduino LEDs, the panel, and the various EL and other lights.



The next video has a room light on and demonstrates some of the control of the robot and screen feedback.



I got some USB speakers too, but they turned out to be a tad too large to mount onto Terry. So I'll get some smaller ones and then Terry can talk to me letting me know what is on its, err, "mind". I guess as autonomy is ramped up it will be useful to know if Terry is planning to navigate around or has noticed that it has been marooned by a chair that a pesky human has moved.

The talk over video is below. I missed talking about the TPLink wifi APs and why there are two, and might be only one in the future. The short answer is that Terry might become a two part robot, with a base station only one wifi AP is needed on the robot itself.

Tiny Tim improves and gets Smaller

I finally switched Tiny Tim over to a lipo battery. Almost everything worked when I tested the new battery, the only thing that failed in a major way were the two 2812 LEDs which, either didn't come on or came on for a very quick moment and went dark. So Tim is now smaller again without the "huge" AA battery pack at it's tail.

The 2812 story was interesting. It wasn't going to be happy jumping to the 7.6v of the 2S lipo. So I tried various voltage divider setups which didn't work either. I ended up using a common 5v regulator and the lights work fine again. I think I was maybe using too high resistor values in the divider and the 2812s didn't like it. At any rate, they apparently want a good regulated power source, and I wasn't giving it one before I switched over to using the regulator.

On the whole, going from 5-6v of the AA pack to 7.6v has made it a snappier mover. I tried it initially with the battery on the bench and found it would lift the back off the desk under hard break.

Next up is probably attaching a claw or drop mechanism and ultrasound sensor and then take on the Sparkfun autonomous ping-pong ball into cup challenge. I'll probably control it via wireless from a second on board micro-controller. The drop, ultrasound, and autonomous navigation micro (and additional battery) can all be put into a single "module" that I can then bolt to Tim. All the navigation micro needs to do is control the differential drive like a remote control would. This way, the existing micro etc on Tim doesn't change at all in order for the challenge to be accepted.

Terry: Updated Top Shelf

The Kinect is now connected much closer to the tilt axis, giving a much better torque to hold ratio from the servo gearbox. I used some self tapping screws to attach the channel to the bottom of the Kinect. Probably not the cleanest solution but it appears to mount solidly and then you get to bolt that channel to the rest of the assembly. For a closer look the Logitech 1080 webcam is mounted offset from the Kinect. This should give an enjoyable time using the 1080 RGB data and combining the VGA depth mask from the Kinect into a point cloud.

The camera pan/tilt is now at the front of the top shelf and a robot arm is mounted at the back of the shelf. The temptation is high to move the arm onto a platform that is mounted using threaded rod to the back of Terry. All sorts of fun and games to be had with automated "pick up" and move tasks! Also handy for some folks who no longer enjoy having to pick items up from the ground. The camera pan/tilt can rotate around to see first hand what the arm is doing, so to speak.

The wheel assembly is one area that I'm fairly happy with. The yumo rotary encoder runs 1024 P/R and it is attached using an 8:1 down ratio to give an effective "ideal world" 13 bit precision. Yes, there are HAL effect ICs that give better precision, though they don't look as cool ;) The shaft of the motor is the axle for the wheel. It is handy that the shaft is not right in the centre of the motor because you can rotate the motor to move the wheel through an arc, and thus adjust the large alloy gear until it nicely mates with the brass gear on the rotary encoder.

Lower down near the wheels is a second distance sensor which is good for up to around 80cm distance. The scan rate is much slower than the Kinect however.

Things are getting very interesting now. A BeagleBone Black, many Atmel 328s on board, and an Intel j1900 motherboard to run the SLAM software.

Terry 2.0 includes ROS!

What started as a little tinker around the edges has resulted in many parts of Terry being updated. The Intel j1900 motherboard is now mounted in the middle of the largest square structure, and SSD is mounted (the OCZ black drive at the bottom), yet another battery is mounted which is a large external laptop supply, the Kinect is now mounted on the pan and tilt mechanism along with the 1080p webcam that was previously there. The BeagleBone Black is moved to its own piece of channel and a breadboard is sunk into the main 2nd top level channel.

I haven't cabled up the j1900 yet. On the SSD is Ubuntu and ROS including a working DSLAM (strangely some fun and games getting that to compile and then to not segv right away).

I used 3 Actobotics Beams, one IIRC is a 7.7 incher and two shorter ones. The long beam actually lines on for the right side of the motherboard that you see in the image. The beam is attached with nylon bolts and has a 6.6mm standoff between the motherboard and the beam to avoid any undesired electrical shorts. With the two shorter beams on the left side of the motherboard it is rather securely attached to Terry now. The little channel you see on the right side up a little from the bottom is there for the 7.7 inch beam to attach to (behind the motherboard) and there is a shorter beam on this side to secure the floating end of the channel to the base channel.

The alloy structure at the top of the pan and tilt now has a Kinect attached. I used a wall mount plastic adaptor which with great luck and convenience the nut traps lined up to the actobotics holes. I have offset the channel like you see so that the center of gravity is closer to directly above the pan and tilt. Perhaps I will have to add some springs to help the tilt servo when it moves the Kinect back too far from the centre point. I am also considering a counter balance weight below the tilt which would also work to try to stabilize the Kinect at the position shown.

I was originally planning to put some gripper on the front of Terry. But now I'm thinking about using the relatively clean back channel to attach a threaded rod and stepper motor so that the gripper can have access to the ground and also table top. Obviously the cameras would have to rotate 180 degrees to be able to see what the gripper was up to. Also for floor pickups the tilt might have to be able to handle a reasonable downward "look" without being too hard on the servo.

There were also some other tweaks. A 6 volt regulator is now inlined into a servo extension cable and the regulator is itself bolted to some of the channel. Nice cooling, and it means that the other end of that servo extension can take something like 7-15v and it will give the servo the 6v it wants. That is actually using the same battery pack as the drive wheels (8xAA).

One thing that might be handy for others who find this post, the BeagleBone Black Case from sparkfun attaches to Actobotics channel fairly easily. I used two cheesehead m3 nylocks and had to force them into the enclosure. The nylocks lined up to the Actobotics channel and so the attachment was very simple. You'll want a "3 big hole" or more bit of channel to attach the enclosure to. I attached it to a 3 bit holer and then attaced that channel to the top of Terry with a few threaded standoffs. Simplifies attach and remove should that ever be desired.

I know I need slip rings for the two USB cameras up top. And for the tilt servo as well. I can't use a USB hub up top because both the USB devices can fairly well saturate a USB 2.0 bus. I use the hardware encoded mjpeg from the webcam which helps bandwidth, but I'm going to give an entire USB 2.0 bus to the Kinect.

Ending up Small

It's easy to get swept up in trying to build a robot that has autonomy and the proximity, environment detection and inference, and feedback mechanisms that go along with that. There's something to be said about the fun of a direct drive robot with just power control and no feedback. So Tiny Tim was born! For reference, his wheels are 4 inches in diameter.

A Uno is used with an analog joystick shield to drive Tim. He is not intended and probably never will be able to move autonomously. Direct command only. Packets are sent over a wireless link from the controller (5v) to Tim (3v3). Onboard Tim is an 8Mhz/3v3 Pro Micro which I got back on Arduino day :) The motors are driven by a 1A Dual TB6612FNG Motor Driver which is operated very much like an L298 dual hbridge (2 direction pins and a PWM). Tim's Pro Micro also talks to an OLED screen and his wireless board is put out behind the battery to try to isolate it a little from the metal. The OLED screen needed 3v3 signals, so Tim became a 3v3 logic robot.

He is missing a hub mount, so the wheel on the left is just sitting on the mini gearmotor's shaft. At the other end of the channel is a Tamiya Omni Wheel which tilts the body slightly forward. I've put the battery at the back to try to make sure he doesn't flip over during hard break.

A custom PCB would remove most of the wires on Tim and be more robust. But most of Tim is currently put together from random bits that were available. The channel and beams should have a warning letting you know what can happen once you bolt a few of them together ;)

3d printed part repo for attaching to Actobotics structure

I started a github repo for 3d print source files to attach things to Actobotics structure. First up is a mount to secure a rotary encoder directly to some channel as shown below.

As noted in the readme, I found a few issues post print, so had to switch to subtractive modelling (dremel time) to complete it. I leant a few things from this though so the issues are less likely to bite me in the future. There are some comments in the scad file so hopefully it will be useful to others and maybe one day I'll get a PR to update it. The README file in the repo will probably be the one true source for update info and standing issues with each scad file.

You can just see the bolt in below the rotary encoder. That has a little inset in the 3d part to stop it from free spinning when you screw in the 6-32 nut from below the channel.

TerryTorial: Wheel Feedback, Object detection, Larger size, and now a screen!

Terry the robot now has grown up a little, both semantically and physically. The 12 inch channel that was the link to the rear swivel wheel now has many friends which are up to 15 inches higher than the base beam. This places Terry's main pan/tilt camera at his top a little bit above table height, so he is far more of a presence than he used to be. On the upside, the wheel controller, shaft encoders, and batteries can all live on the lower level and there is more room for the actual core of Terry up top. I'm likely to add one or two more shelves to the middle of Terry for more controllers etc.

There is now also a 16x32 RGB matrix screen up front so Terry can tell you what he is thinking, what speed his wheels are rotating at, and if there is an obstacle that has been detected.

I'm rendering the framebuffer for the screen on the BeagleBone Black using Cairo and Pango (thus & freetype). The image data is extracted from Cairo as 32bit RGBA and packed down into planar data that the arduino expects. That packed framebuffer is then sent over UART to an Arduino 328 which takes care of refreshing the RGB matrix so that fake colour levels are achieved using a software PWM/BAM implementation. Yay for the BBB having 4.5 UARTs. I need to work out how to bring up the TX only UART as that is perfect for the one way communication used to drive the screen.

I should be able to get better colour precision using a Teensy 3.x as the screen driver. The Cortex-M4 just has more cycles to be able to softpwm the screen faster to be able to get a greater perceived colour range.

I customized the font a little in FontForge. Specifically I modified the kerning for "Te" to not waste as much space. The font is based on Cantarell by Dave Crossland. I've only just started on this open font to LED matrix stuff, but with some form of PWM and FreeType rendering the data I hope to be able to get nice antialiased font renders, even at the low resolution of 16x32.

The wheel encoders make a huge difference to the whole experience. Even without full autonomy the encoders allow you to have a repeatable oval or patrol path which can be followed. Also being able to setup the speed to avoid large acceleration or jerky movement so that the higher Terry is still a very stable Terry when on the move.

To beep or not to beep.

There comes a time in the life of any semi autonomous robot when the study of the classics occurs. For Terry, the BeagleBone Black driven Actobotics construction, short of visiting the Globe the answer would seem to be at current "To Beep".

Terry is now run by a RoboClaw 2x5A controller board with two 1024 pulse per revolution shaft encoders attached to an 8:1 ratio gearing on the wheel. This gives an overall of about 13 bits of precision from the shaft encoders relative to the wheel. In the future I can use this magnificent 4 inch alloy gear wheel as a torque multiplier by moving the motor to its own tooth pinion gear. Terry now also has an IR distance sensor out front, and given the 512mb of RAM on the BeagleBone some fairly interesting mapping can be built up by just driving around and storing readings. Distance + accurate shaft encoders for the win?

To use the RoboClaw from the BeagleBone Black I created a little nodejs class. It's not fully complete yet, but it is already somewhat useful. I found creating the nodejs interesting because when coding for MCUs one gets quite used to the low level style of sending bytes and stalling until the reply comes in. For nodejs such a style doesn't implement well. Normally one would see something like

doMagic( 7, rabbits, function() {
    console.log("got them rabbitses, we did");
});

Where the callback function is performed when the task of pulling the number of rabbits from the hat is complete. For using the RoboClaw you might have something like

claw.getVersion( function( v ) {
  console.log("version:" + v.value );
});
claw.setQPPS( 1024 );

Which all seems fairly innocent. The call to getVersion will send the VERSION command to the roboclaw with a given address (you can have many claws on the same bus). The claw should then give back a version string and a checksum byte. The checksum can be stripped off and verified by the nodejs.

The trouble is that the second call, to set of the quadratic decoder pulses per second, will start to happen before the RoboClaw got a chance to service the getVersion call. Trying to write a second command before you have the full reply from the first is a recipe for disaster. So the nodejs RoboClaw class maintains a queue of requests. The setQPPS() is not run right away, but enqueued for later. Once the bytes come back over the UART in response to the getVersion() then the callback is run and the RoboClaw nodejs class then picks the next command (setQPPS) and sends that to the RoboClaw. This way you get the strict ordered serial IO that the RoboClaw hardware is needing, but the nodejs can execute many commands without any stalling. Each command is handled in the order it is written in the nodejs source code, it just doesn't execute right away.

Next up is trying to tune the PID controller variables for the robot. One might expect a command such as "move 1 wheel circumference forward" to work fairly exactly if the quality of the shaft encoders is good. Though for the wheel size and QPPS I get a little bit of overshoot currently. I suspect the current D is not sufficient to pull it up in time. It may be a non issue when the wheels are down and friction helps the slow down process.

Terry now also has an IMU onboard. I created a small class to read from the TWI HMC5883 magneto. You'll have to figure your declination and I have a small hack on about line 100 to alter the heading so that a 0 degree reading means that physically Terry is facing North. I should bring that parameter out of the class so that its easier to adjust for other robots that might mount their IMU in a different physically orientation than what I happened to use.