Paper Review 31-10-2018
TERMES: An Autonomous Robotics System for Three-Dimensional Collective Construction
K. Petersen, R.Nagpal, J.Werfel, SEAS/Wyss Institute, Harvard University, Robotics: Science and Systems Conference 2011
Aims of the paper
The paper presents a first step towards construction with swarms of robots using specially designed robots and construction materials. It primarily conveys the reasoning for the design of elements of the swarm construction system including climbing, locomotion, navigation and manufacturing. It appears to be aimed at notifying the research community that such a platform exists with a lot of opportunity and scope for further research in a wide variety of fields.
The author’s stated goal of their system is to design an autonomous robot which can collect building material and use it to build a user specified target structure on a flat, obstacle free workspace. The approach at engineering a solution inspired by the decentralized construction performed by termite colonies. This is a novel solution as previous ideas have only been theoretical, or in 2D applications with few being physical realizable due to issues with reliability of the attachment mechanisms for 3D systems. This particular solution comprises of 2 parts:
- A Specialized passive building blocks which can be lined up in a grid as well as stacked for height.
- A series of identical robots which can autonomously manipulate these blocks into structures. These robots have the ability to move across, climb and turn about these blocks to build much larger structures than themselves.
This paper outlines the decision process of designing the robot to reliably solve the problems of climbing to reach higher levels, navigation to move safely and locate itself within the structure and transporting blocks. In each section the author conducts an experiment to prove the reliability of their chosen method.
Climbing This task focuses on the method of locomotion to allow both movement and climbing of blocks. From experimentation, whegs (wheeled legs) were chosen as the best method on comparison of maximum climbing height, self-alignment and gait smoothness. The blocks were also modified by adding notches to give the robot’s whegs steps to climb taller blocks and correct for misalignments.
Navigation was solved by printing a white cross on the blocks and using several infrared (IR) sensors on the bottom of the robot. Progress can then be monitored by a known set of IR observations. Testing the technique by asking the robot to repeatedly traverse a path 40 times was successful with no errors.
Transportation was solved by designing a small singularly actuated arm which could attach to specifically machined slots in each block. The arm’s forks are controlled by springs only. A series of pushbutton sensors register when the forks are down, up, or carrying something.
The robot is then simply programmed using a Logo turtle like manner. A simple pseudocode is provided which implements some simple collaboration but it is not the focus of this paper and is not implemented in reality. It is worth noting for the examples, the initial blocks, as well as the foundation are already given by the user before the robot starts building. The robot and blocks are mostly made of off the shelf components with the shell, arms and claw being 3D printed and the axle which is machined. It was designed to be easily mass-built and replicable.The robot itself is controlled by an ATmega1281 and has sensor processing hardware, motor drivers IR and tilt sensors (for navigation) and also Bluetooth for communication with a PC. The Blocks are fabricated out of foam and each have magnets in the corners to aid in the construction process. The author concludes by saying that this is just the initial platform and further hardware and software improvements will be necessary for fully autonomous construction.
Coming from a software background, it was an interesting change of style to the papers I am most used to reading. This paper was really well structured and the experimentation, observation, decision pattern is much clearer than some of the software papers I’ve read.
Examining their content, at each section we see criterion which logically follow from the problem they wish to solve.
In climbing, the 3 criterion chosen do indeed appear to be complete enough for their task. In experimentation, they were very thorough about keeping the conditions similar and fairly tested a large variety of locomotion options. Perhaps more trails could have been done but the results presented seem sufficient and would follow intuition. Another suggestion is I would imagine that each of the locomotion configurations may be able to perform just as well if they had custom designed both the robot and the blocks around it just like they’ve done currently with the whegs. However it is to be said that no experimental evidence was given to some of the smaller design choices such as the curving of the whegs and the notches even if they seem like good ideas. Further to this, another justification for whegs follows from travel times across unstructured terrain, even though no other locomotion methods were tested.
In navigation, the chosen technique feels rather rudimentary, but, according to the results, rather effective. It would have been interesting to see if other techniques for navigation were tested as only the one used is presented. Also it is worth noting that the blocks themselves are purposely designed and because the author controls the whole system, a method such as this is possible - it cannot be generalised (although this is not the aim of this work). Perhaps this is why the repeated traversal experiment is so successful with surprisingly no errors. This technique could definitely be applied in other closed systems were reliability is especially important. I would like to see it in action for larger perhaps more complex structures and also failure cases. This experiment can be seen to be playing into the strengths of the bot, but what would happen if a block was suddenly removed, or the robot suddenly found a stack that was too high to traverse. These cases may stretch the initial assumptions but for reliability would be good to observe.
In manipulation we observe similar constraining of the system within the experimentation and the fact that only the used method is detailed with no alternatives. I do, however, like the simplicity of the solution and how careful thought has been made to reduce actuation to the minimum.
Overall this was a good paper which was relatively thorough in its analysis of its own designs. It referenced plenty of old attempts, but it would seem that its approaches are entirely novel with no literature cited as inspiration for its designs (except for the use of whegs). I believe that the author has indeed met their initial aims of introducing the system and showing how it is both a reliable and robust design that can indeed complete rudimentary programmable construction tasks.
I would recommend a full reading of this paper for researchers interested in both swarm robot design, and swarm control as relevant background reading. Following this paper I recommend reading the follow up paper by the same authors which focusses on distributed multi-robot algorithms specifically for the TERMES system.