Designing Robot Swarms and Bio-hybrid Systems for Adaptivity and Robustness

Heiko Hamann
Institute of Computer Engineering, University of Lübeck, Germany
hamann@iti.uni-luebeck.de

Decentralized self-organizing large-scale systems, such vironment (location and intensity of light spots changes). In
as robot swarms (Hamann, 2018), can be designed to be an additional experiment, the robot swarm is halved and re-
adaptive, robust, and scalable. However, developing robot duced to swarm size N = 5 during the experiment. For both
and system behaviors that are robust and adaptive to dy- adaptation to dynamic environments and adaptation to dy-
namic environments, dynamic system size, and faults is still namic system size, the robots face an interesting and funda-
challenging. We study a swarm showing robust scalabil- mental challenge. The adaptation can either be fast or accu-
ity. The swarm needs to detect the change in its environ- rate. A fast adaptation needs to be sensitive to any detected
ment or in the system size, assess the quantity of the change, change that needs to be considered a precursor of changes
and appropriately adapt parameters of its control algorithm. requiring adaptations in the robot behavior. However, if
In a second part, we discuss bio-hybrid systems of natural we require a robust adaptation process, then robots need to
plants interacting with robotic nodes. In the studied exam- avoid false positives (reacting to a change where there was
ple, robotic devices are used to steer the growth of natural no change). This is a tradeoff where improving in one ca-
plants by exploiting their adaptive behaviors. We use meth- pability results in worsening the other. Wahby et al. (2019)
ods of machine learning to model natural plants and to guide resolve that challenge by periodical measurements of envi-
their growth and motion with an autonomous system. In bio- ronmental features (e.g., light) and other features (e.g., times
hybrid systems, we can exploit natural adaptive behaviors to between robot-robot encounters) indicating swarm density.
build, for example, systems that self-repair. These measurements are averaged over a limited time win-
dow and directly influence control parameters of the robot.
Adaptivity in Swarm Robotics Hence, the robots do not explicitly react to a detected change
but forget old measurements that were recorded before the
Swarm robotics is claimed to have the advantage of high change. Besides measuring light intensities and remember-
robustness against failures (Brambilla et al., 2013). Their ing maximum/minimum light intensities, robots measure the
high degree of redundancy helps to overcome failures, such time ta between two encounters of a wall and the time tr be-
as breaking robots. Similarly, swarm robot behaviors are tween two encounters of a robot. If we assume a regular
claimed to be scalable to different swarm sizes and some- floor plan (e.g., rectangular) without many obstacles (except
times also to different swarm densities (robots per area). for other robots), then the time between encountering walls
High degrees of scalabilty can be achieved as swarms rely indicates the side length and hence the area of the room. The
on local information and local communication only and time between encountering other robots is similar to a mean
has been shown in experiments (Rubenstein et al., 2014; free path and indicates the robot density. The distribution of
Hamann, 2018). We also know that optimal swarm densities measured times between robot-robot encounters is similar to
exist for many systems of swarm robotics (Hamann, 2013). an exponential distribution and its mean, for example, drops
However, if robots break at runtime, then the swarm den- significantly if the robot density is halved. Robots meeting
sity changes and the swarm possibly requires online scala- robots on a light spot stay stopped for waiting time w. The
bility to remain efficient. We require the swarm to be robust time w is a parameter of the control algorithm that depends
against dynamic swarm sizes which can be called ‘robust on these measured values. w is scaled proportionally to the
scalability.’ If the swarm size changes, then each robot may measured time tr . Measured time ta was not used here as
require to change its behavior at runtime, for example, by the area remained constant in this experiment. The results
adapting parameters of its control algorithm. indicated that robots adapt successfully to a change of sys-
Wahby et al. (2019) present an aggregation experiment tem size from N = 10 to N = 5 at runtime and outperform
with N = 10 swarm robots. The robots’ task is to aggregate a swarm that was optimized offline for N = 10.
at light spots and each robot needs to adapt to a dynamic en-

Copyright c 2021 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
 Robustness in a Bio-hybrid System defined pattern turns on its blue LEDs to attract the plants.
Bio-hybrid systems combine living organisms with tech- Once α detects the first close-by plant, α communicates with
nology. Here, we combine natural plants with robot-like the next robot β of the programmed pattern. α turns off its
units. Key of the bio-hybrid approach is that we get ‘life- blue LED and β turns its blue LED on. This process repeats
like’ features almost for free instead of trying to mimic be- until the plant approaches the last robot in the sequence and
haviors of living organisms in a purely technological ap- the pattern has been grown. An experiment takes several
proach. An advantage is, for example, that a bridge built weeks and was repeated successfully.
from living plants reinforces itself and grows stronger over In future work, we plan to show a robust bio-hybrid sys-
time instead of decaying slowly but inevitably (Shankar, tem with the capability of self-repair. We want to scale up
2015). A challenge is that we need to understand how to one more step (dozens of plants and 18 robotic devices) and
interface the plants and how to exploit their natural adap- to use a more complex scaffold. After having grown the
tive behaviors. In the EU-funded project flora robotica, plants for several weeks, we plan to punch a whole into the
we have developed a bio-hybrid system of living organisms scaffold. The system is then expected to regrow that part
and robotic devices to steer and guide the growth and mo- while keeping other areas (e.g., windows) unobstructed.
tion of natural plants (Hamann et al., 2015). We use the
References
phototropism (growth towards light) and the thigmotropism
Brambilla, M., Ferrante, E., Birattari, M., and Dorigo, M. (2013).
(growth guided by touch, for example, in climbing plants) of Swarm robotics: a review from the swarm engineering per-
plants. Bright blue LEDs are used to attract plants and scaf- spective. Swarm Intelligence, 7(1):1–41.
folds can be used to determine the growth options of climb-
Hamann, H. (2013). Towards swarm calculus: Urn models of col-
ing plants (Wahby et al., 2018a). Here, we shortly discuss lective decisions and universal properties of swarm perfor-
two experiments: (A) high-precision control of growth and mance. Swarm Intelligence, 7(2-3):145–172.
motion of a single plant using a single robotic device and
(B) growth of a pattern by guiding a small group of plants Hamann, H. (2018). Swarm Robotics: A Formal Approach.
Springer.
using multiple robotic devices.
In experiment A, we use a simple setup with one plant Hamann, H., Wahby, M., Schmickl, T., Zahadat, P., Hofstadler,
(common bean), two lights (left/right), and a camera. We al- D., Støy, K., Risi, S., Faina, A., Veenstra, F., Kernbach, S.,
Kuksin, I., Kernbach, O., Ayres, P., and Wojtaszek, P. (2015).
low a user to define target points in 2-d space that the plant’s flora robotica – mixed societies of symbiotic robot-plant bio-
tip should visit during the experiment (Wahby et al., 2018b). hybrids. In Proc. of IEEE Symposium on Computational In-
Our tool-chain to solve this engineering task is rather com- telligence (IEEE SSCI 2015), pages 1102–1109. IEEE.
plex. We start from a dataset obtained by preliminary exper- Jakobi, N., Husbands, P., and Harvey, I. (1995). Noise and the
iments where the two lights are switched on/off in a regu- reality gap: The use of simulation in evolutionary robotics.
lar sequence. The plant is photographed every five minutes. In Proc. of the 3rd European Conf. on Advances in Artificial
Using computer vision, we extract data that represents the Life, volume 929 of LNCS, pages 704–720. Springer.
plant’s reaction to given light conditions and a previous stem Rubenstein, M., Cornejo, A., and Nagpal, R. (2014). Pro-
configuration (e.g., length, bend). With that data we train an grammable self-assembly in a thousand-robot swarm. Sci-
LSTM network to obtain a holistic plant model that predicts ence, 345(6198):795–799.
the plant’s reaction for a given configuration. We use the Shankar, S. (2015). Living root bridges: State of knowledge, fun-
LSTM network and methods of evolutionary computation to damental research and future application. In IABSE Sympo-
evolve a controller that takes the plant’s configuration, the sium Report: Structural Engineering: Providing Solutions to
light condition, and the user-defined target points as input Global Challenges, volume 1, pages 1974–1981.
and outputs the desired next configuration of the lights. In a Wahby, M., Heinrich, M. K., Hofstadler, D. N., Neufeld, E.,
last step, we use the light controllers that performed well in Kuksin, I., Zahadat, P., Schmickl, T., Ayres, P., and Hamann,
simulation to control a real plant and find that they succeed H. (2018a). Autonomously shaping natural climbing plants:
despite an expected reality gap (Jakobi et al., 1995). A bio-hybrid approach. Royal Society Open Science,
5:180296.
In experiment B, we support the growth of a group of bean
plants with a scaffold in the form of a diagrid. In the bifurca- Wahby, M., Heinrich, M. K., Hofstadtler, D. N., Zahadat, P., Risi,
tion points we place eight robotic devices that use proximity S., Ayres, P., Schmickl, T., and Hamann, H. (2018b). A robot
to shape your natural plant: The machine learning approach
sensing to detect a close-by plant and that have red and blue
to model and control bio-hybrid systems. In GECCO, pages
bright LEDs (Wahby et al., 2018a). These robotic devices 165–172.
can also communicate between each other using WLAN.
The task is to grow a user-defined pattern along the dia- Wahby, M., Petzold, J., Eschke, C., Schmickl, T., and Hamann, H.
(2019). Collective change detection: Adaptivity to dynamic
grid. The pattern is defined and programmed into the robots. swarm densities and light conditions in robot swarms. In
After planting several bean plants, the autonomous system Artificial Life Conference Proceedings, pages 642–649. MIT
controls the remaining process. A first robot α in the user- Press.