Day 8 - Robotics with neuromorphic brains (Matej Hoffmann, Jing Shuang (Lisa) Li, Helmut Hauser, Guido De Croon)
Robotics with neuromorphic brains
(Tobi Delbruck, solely responsible for inaccuracies)
Elisa Donati MC'ed this session to keep speakers to their allotted time.
Matej Hoffmann (Czech Technical U in Prague) started things off. He came from Rolf Pfeiffer's lab in Zurich which focused for many decades on embodied robotics, where some of the functionality is embedded in the natural physics of the robot coupling to the environment, for example the swinging of legs arises from their mass and joints and inertia by sometimes including springs to recover energy.
Famous examples are passive walkers and dead trout that swims upstream.
For decades, walking robots had a kind of Parkinson's victim bent-knee stance to increase stability to disturbances but this resulted in truly awful cost of transport (a dimensionless measure of cost of moving).
Recent bipedal walkers go in the direction of incorporating efficient passive dynamics, from Boston Dynamics and others, including a Chinese company whose biped completed a half marathon on one charge.
Matej then showed this table to compare neuromorphic and conventional robotics:
Jean-Jacques Slotine (MIT) pointed out that NNs are used for adaptive robot control in MP robots routinely now.
In relation to SOA walking and flying robots, since they now use small NNs for most of the low level control, the power consumption is a minor fraction of the motor power, in contrast to past where solving MPC dominated and comprised about half the power.
Robustness: Sim2Real training with domain randomization coupled with some control theory can make robots robust for some tasks, e.g. balancing. JJ pointed out that biology manages high robustness even with huge perception action latencies. BW pointed out that all SOA robotics benefits immensely from using ML approaches. GC pointed out that NN controlled robotics are more interpretable than biology bccause all state is measurable. MH pointed that SOA robotics needs immense amount of (mostly simulated) data. I pointed out that domotics are better (robot vacuums) but only affciandos tolerate the TLC they require.
Lisa Li (U Michigan) went on from there. Her expertise of control theory, which must guarantee stability under worst case disturbances. She drew a diagram of the modular arrangement of robots that allows their design.
image TBA
She then wrote this set of guidelines for design of NM robots:
image TBA
The main conclusion is that NM or NN controllers are most appropriate for "throwaway" robotics that does not need to guarantee robustness under worst-case disturbances or certain control performance characteristics.
Lisa kindly supplied this list of further reading:
- Example of modularity/architecture of current robots (focusing on the ways hardware/software typically interface):
- Kortenkamp et al. -- Robotic systems architectures and programming
- https://link.springer.com/chapter/10.1007/978-3-319-32552-1_12
- This is part of the "Springer Handbook of Robotics"; looking through the chapter titles, particularly of parts A-C, also gives an idea of how people think about and separate "modules" of robots
- Examples of neuromorphic computing/hardware for robotics:
- Yang et al. -- Neuromorphic electronics for robotic perception, navigation and control: A survey
- https://www.sciencedirect.com/science/article/pii/S0952197623010229
- Examples of how existing learning-based methods can integrated into robotics while preserving safety:
- Brunke et al. -- Safe Learning in Robotics: From Learning-Based Control to Safe Reinforcement Learning
- https://www.annualreviews.org/content/journals/10.1146/annurev-control-042920-020211
After the coffee break, Helmut Hauser (Bristol U) took over to discuss soft robotics (SR), in particular morphological computation. Soft robots use soft components, actuators, grippers, etc. They are harder to control because of the huge compliance and complex dynamics, particularly in open-world scenarios.
Soft robotics is an emerging field, with potential application in fruit picking, exoskeletons, domotics (domestic robots), and handling of delicate materials. People cans till publish very speculative papers about components without clear path to practicality since the boundaries are still being established.
Morphological robotics is still arranged in the hierarchy above. The body is now very flexible and soft and very hard to model with conventional ODE based methods. Babies provide some inspiration when they first start to motor-babble to learn their own body dynamics.
After a discussion of what is SR, Helmut went on to two things where NM might be applicable.
1. Can we have morphology that "spikes"? Can we build e.g. a tactile sensor that turns rubbing over a surface to oscillations that turn into change events? Helmut pointed out that the hairs need to have diverse lengths to provide non-redundant information, but that such tactile sensor is ideal use case for activity-driven event-based sensing. This area is being investigated in other labs currently, notably Chiara Bartolozzi's lab at IIT Genova.
Barbara Webb pointed out the example of a filament from light bulb (which is highly resonant) that makes intermittent contact during movement. Another example I pointed out is the biobug from Mark Tilden and Wowwee robotics, where the feelers were designed to resonant with the walking and could be used to sense forward walking progress on the 4-bit microcontroller with 11 instructions that could count the contacts of the feeler with the metal ring at its base.
2. Can we make a multilayered neural network that uses the body as the earliest feature detector? He drew this sketch
It shows a lovely example of an octopus arm in water that used the arm joints as the resevoir units, in a feedback loop for arm control. The memory is the state of the arm. The only learning is the training of the readout. SK asked how is this not just a linear controller? (since the readout is linear weighted sum of robot joint states.)
Guido de Croon asked about epistatic actions, which are just taken to gather information (like looking around). MH agreed that a limitation of this approach is that it is harder to include these epistatic actions since it based on supervised training and that anything outside this experience will fail.
.
3. The thing that Helmut is exploring now is adaptive SR, where for example skin that is added to bodies, and where the skin stretches, e.g. on joints, is informative compared with other areas.
Helmut kindly supplied an extensive list of further material (edited to be a bit shorter for the blog)
Theoretical Models
Here are our two theoretical models on morphological computation.
The underlying idea is to apply the concept of reservoir computing to
networks of nonlinear mass-spring-damper systems, which are a good way
to describe biological and soft bodies. Note
that they are a dense read (below are more high-level papers). However,
for the mathematically minded, it might be joyful.
Hauser, H.; Ijspeert, A.; Füchslin, R.; Pfeifer, R. & Maass, W.
"Towards a theoretical foundation for morphological computation with compliant bodies"
Biological Cybernetics, Springer Berlin / Heidelberg, 2011, 105, 355-370
Hauser, H.; Ijspeert, A.; Füchslin, R.; Pfeifer, R. & Maass, W.
"The role of feedback in morphological computation with compliant bodies"
Biological Cybernetics, Springer Berlin / Heidelberg, 2012, 106, 595-613
Morphological Computation for Control
This is part of a special on Control for Soft Robots. In our
contribution, we discuss different ways in which the morphology could
help in the control of the robot.
The paper is written to serve as a resource of inspiration. It discusses research opportunities in this space.
Leveraging morphological computation for controlling soft robots: Learning from nature to control soft robots
H Hauser, T Nanayakkara, F Forni
IEEE Control Systems Magazine 43 (3), 114-129
Implications for robotics
This is a more high-level paper where I try to explain the insights
gained from the two theory papers. I discuss ideas on a much higher
level. The idea was to make it more accessible to people coming from
other fields.
Helmut Hauser
”Physical Reservoir Computing in Robotics”
Using it with real robots
Here are a few papers where we used real robots as reservoirs (there are more)
Nakajima, K.; Hauser H.; Li T.; and Pfeifer R.
"Information processing via physical soft body"
Scientific Reports 5 (2015) Article number: 10487,
doi:10.1038/srep10487
Spine dynamics as a computational resource in spine-driven quadruped locomotion
Q Zhao, K Nakajima, H Sumioka, H Hauser, R Pfeifer
2013 IEEE/RSJ International conference on intelligent robots and systems
Tapered whisker reservoir computing for real-time terrain identification-based navigation
Z Yu, SMH Sadati, S Perera, H Hauser, PRN Childs, T Nanayakkara
Scientific Reports 13 (1), 5213
(Not technical on a real robot, but more realistic simulation)
Nakajima K.; Li T.; Hauser H.; and Pfeifer R.
"Exploiting short-term memory in soft body dynamics as a computational resource"
Journal Royal Society Interface, 6 November, 2014, vol. 11, no. 100, 20140437
DOI: 10.1098/ rsif.2014.0437
Philosophical Aspects of Morphological Computation
This discusses some philosophical aspects of morphological computation.
The morphological paradigm in robotics
S Freyberg, H Hauser
Studies in History and Philosophy of Science 100, 1-11
Finally Guido de Croon took over to discuss the future of autonomous robotics and asserted that it will be increasingly neuromorphic.
His group works on small agile drones. After his PhD, he decided to work on tiny robots that need to work with tiny amounts of power and weight. The DelFly is a flapping wing flyer, The wings in the original DelFly just provided thrust and steered by using a rudder. The new one, Agile DelFly has two motors and a servo. The original weighed 16g with 22cm and 1-3W and about 30m fast flight time, the new one 30g and 32cm wingspan with about 3W power and only 5m flight time. The battery weighs about 5g. The original could hover only in hands of expert but could do stable forward flight, The new one uses an IMU to measure acceleration (on good day) and gyro. The new one needs to control at about 1kHz. These microflyers cannot possibly carry an NVIDIA Jetson that weighs at least 80g and burns >12W.
Drone weigh about 1kg and burn about 200-300W to fly. There is a slow boat effect(?) that was not clear to me.
Drones need to control yaw pitch and roll. Also they need to control rate of turning. Typical autopilots help with attitude control but inctreasingly assist with other tasks such as turn rate and other maneuver control including open-air waypoint navigation, though they are not yet capable of robust obstacle avoidance. He then sketched the typical drone control hierarchy and their update charactertics. They are working on SNN formulations of various parts of this stack.
They recently won in a drone race at Abu Dhabi reaching up to 100km/h using RL based AI control. In the (conventional camera) images you see a gate. Their CNN highlights the gates and segments the gates with about 1M weights.
He concluded by describing their success at the drone racing competition and kindly supplied this set of further reading
- Fully neuromorphic ego-motion estimation and control
- Neuromorphic attitude estimation and control
- Drone racing
- https://youtu.be/yz2in2eFATE?si=pEXAwDRyqvVhL5su
- a good indication for what we do in control for the drone race
- - a good indication of what we do in perception for the drone race
That's was it for a really interesting session.
The afternoon was filled by workgroup meetings and the very last attempt to sail Argo, my robot sailboat. Once again Argo is a bit too leaky and the seawater killed my RPi, literally burning out the power regulators as you can see in the image below :-(. Next year I will come back with a truly watertight boat, waterproofed electronics, and a proper PCB design that can be quickly replaced.
However I this very last run--where I had to swim out to rescue the boat:
.jpg)
did result in real data:
That's it for day 8, see you tomorrow!
Comments
Post a Comment