There was a scarcity of entries within the tablebot competitors shortly earlier than the registration window closed for RoboGames 2023. To make certain the competition can be held, I entered a robotic. Then I needed to construct one.
What’s a tablebot?
A tablebot lives on the desk. There are three “phases” to the competitors:
- Phase I: Build a robotic that goes from one finish of a desk to the opposite and again.
- Phase II: Have the robotic push a block off the ledge of the desk.
- Phase III: Have the robotic push the block right into a shoebox mounted on the finish of the desk.
There can be an unofficial Phase IV – which is to fall off the desk and survive. I didn’t try this part.
The majority of tablebots are fairly easy – a few sonar or IR sensors they usually sort of wander across the tabletop in hopes of finishing the completely different phases. My tablebot is decidedly completely different – and it paid off because the robotic received the gold medal at RoboGames 2023.
Robot construct
The complete robotic is constructed of 3D printed components and random issues I had readily available.
I’ve had a type of $99 LD-06 lidars sitting round for some time, and determined this was an awesome challenge to apply it to. I used a Dynamixel AX-12 servo to tilt the laser so I can discover the desk, the dice, or the purpose.
All of the code runs on an STM32, on my customized Etherbotix board which was designed for my Maxwell robotic quite a few years in the past. The robotic makes use of differential drive with some 30:1 12V gear motors, which have been bought from Lynxmotion in 2008 and utilized in numerous hearth combating robots over time.
A set of small digital Sharp IR sensors are used as cliff sensors. These may be moved up or right down to calibrate for various desk surfaces utilizing a pair of adjustment screws. While the sensors are very correct and cease the robotic, they don’t see far sufficient forward when going at full pace, and so I additionally use the laser to detect when the desk edge is approaching.
Phase 1 Software
Phase 1 is fairly straight ahead – and largely primarily based on useless reckoning odometry:
- The laser is angled downwards in search of the desk. This is completed by projecting to the scan to 3D factors, and filtering out something not in entrance of the robotic at roughly desk top. When the desk disappears (variety of factors drops too low), we cut back our most pace to one thing that’s secure for the cliff sensors to detect.
- While the laser sensors search for the top of the desk, the robotic drives ahead, and a easy suggestions loop retains the robotic centered on the desk utilizing odometry.
- When the cliff sensors finally set off, the robotic stops, backs up 15 centimeters, after which turns 180 levels – all utilizing useless reckoning odometry.
- The most pace is then reset and we take off to the opposite finish of the desk with the identical conduct.
Phase 2 Software
The actions of Phase 2 are mainly the identical as Phase 1 – we drive ahead, staying centered with odometry. The pace is a bit decrease than Phase 1 as a result of the laser can be in search of the block:
- The laser scan is projected to 3D, and we filter out any factors which are a part of the desk primarily based on top. These remaining factors are then clustered and the clusters are analyzed for dimension.
- If a cluster is an efficient candidate for the block, the robotic flip in the direction of the block (utilizing, you guessed it, useless reckoning from odometry).
- The robotic then drives in the direction of the block utilizing a easy management loop to maintain the heading.
- Once the block is arrived at, the robotic drives straight till a cliff sensor journeys.
- At that time, the robotic stops the wheel on the aspect of the tripped cliff sensor and drives the opposite wheel very slowly ahead in order that we align the entrance of the robotic with the sting of the desk – making certain the block has been pushed off the desk.
Phase 3 Software
The last part is probably the most advanced, however not by a lot. As with the sooner phases, the robotic strikes down the desk discovering the block:
- Unlike in Phase 2, the robotic really approaches a pose simply behind the block.
- Once that pose has been reached, the robotic tilts the laser again to stage and finds the purpose.
- The robotic then turns in the direction of the purpose in the identical manner it first turned in the direction of the block.
- The robotic then approaches the purpose utilizing the identical easy management loop, and within the course of finally ends up pushing the block to the purpose.
All of the software program for my Tablebot is availble on GitHub.
Robogames video
Jim Dinunzio, a member of the Homebrew Robotics Club, took a video in the course of the precise competitors at Robogames so you possibly can really see the profitable set of runs:
Visualization
To make improvement simpler, I additionally wrote a Python GUI that renders the desk, the robotic odometry path, the laser knowledge, and detected targets and cubes.
Fun with math
Along the best way I really ran right into a bug within the ARM CMSIS DSP library. I used the arm_sin_cos_f32() perform to compute my odometry:
arm_sin_cos_f32(system_state.pose_th * 57.2958f, &sin_th, &cos_th);
system_state.pose_x += cos_th * d;
system_state.pose_y += sin_th * d;
system_state.pose_th = angle_wrap(system_state.pose_th + dth);
This perform takes the angle (in levels!) and returns the sine and cosine of the angle utilizing a lookup desk and a few attention-grabbing interpolation. With the visualization of the robotic path, I seen the robotic odometry would sometimes bounce to the aspect and backwards – which made no sense.
Further investigation confirmed that for very small unfavourable angles, arm_sin_cos_f32 returned enormous values. I dug deeper into the code and located that there are a number of completely different variations on the market:
- The model from my older STM32 library, had this explicit concern at very small unfavourable numbers. The similar bug was nonetheless current within the official CMSIS-DSP on the arm account.
- The model within the present STM32 library had a repair for this spot – however that repair then broke the perform for a whole quadrant!
The concern turned out to be fairly easy:
- The code makes use of a 512 ingredient lookup desk.
- For a given angle, it has to interpolate between the earlier and subsequent entry within the desk.
- If your angle fell between the 511th entry and the subsequent (which might be the 0th entry resulting from wrap round) then you definately used a random worth within the subsequent reminiscence slot to interpolate between (and to compute the interpolation). At one level, this resulted in sin(-1/512) returning outrageous values of like 30.
With that bug mounted, odometry labored flawlessly afterwards. As it turned out, I had this similar perform/bug current in some brushless motor management code at work.
Robogames wrap up
It is superior that RoboGames is again! This little robotic received’t be making one other look, however I’m beginning to work on a RoboMagellan robotic for subsequent yr.
Michael Ferguson