Fracción I. Comisiones del ayuntamiento
EXPEDIENTE: SHA/099/CABILDO/
IV.- Con fecha 24 de agosto del 2017, reunidos en Pleno los Integrantes de la Comisión de Hacienda, aprobaron el presente Dictamen
The MISO convolution model was effective for understanding how to predict of
propulsive force from a compliant, multiple degree of freedom robotic fin with complex
dynamics. It was effective for prediction of forces from sensory data and provided insight
into the underlying relationships between sensor parameters and forces. The MISO model
structure allowed for an in depth analysis of the roles of individual sensors, the memory
of the system, and the performance effects of changing multiple parameters. The
underlying model could be further improved using basic knowledge of robot behavior
(e.g. which gait and stroke phase are being executed). The evaluation methods used,
including MISO for prediction and AIC for sensor and parameter selection, suggest a
framework with which to optimize sensor selection and placement for intrinsic robotic
research areas, including: multi-modal tactile sensing in perception research [84, 115],
grasping studies of in-hand manipulation [116, 117], and proprioceptive robotic systems
for measurement of human body forces and kinematics [118]. The empirical techniques
presented here could be used to validate optimal sensor modalities and optimal sensor
placements for many other sensing challenges in robotics, especially where the dynamics
of the system are challenging to model mathematically but feasible to model with bio-
inspired robotic platforms [119]. These techniques could be valuable to the study of
animal systems, where compliance, complex dynamics, and environmental effects all
contribute significantly to the performance of particular behaviors.
Complex changes to the fin kinematics required more complex sensing and
modeling strategies. When small changes were made to the fin dynamics, by changing
flapping frequency or fin stiffness, the same model could be used to predict forces across
these conditions (Figure 39). But when the entire swimming gait changed, the model
trained on one gait could not be used to predict the forces of another gait. For instance,
models trained on steady swimming forces were unsuccessful at predicting the forces of
ventral steady swimming (Figure 42), even though ventral steady swimming has the same
ventral edge kinematics as steady swimming (Figure 37). Sensing strategies also had to
change as gaits changed. The best sensors subsets for each gait differed, as best
predictions of ventral steady swimming forces were achieved with S={C,CO,A,B} and best predictions of steady swimming forces were achieved with S={BO,A}. The fluid structure interaction that produces propulsive forces involves multiple fin regions and the
energetic exchange of their interactions [70, 120], and so large changes to kinematics can
sensory system that can reliably predict forces when swimming gaits change, multiple
sensors and models are needed. This may have implications for robots that execute
complex gaits and gait changes with compliant control surfaces, such as elastomer-based
soft robots that crawl and undulate [121], fish-inspired robots that use compliant fins to
execute multiple swimming gaits [33, 109, 122], and bipedal robots with compliant joints
that shift from walking to jogging [123]. Robots that engage in multi-modal locomotion
(e.g. aquatic to terrestrial [124, 125], terrestrial to aerial [126], aerial to scansorial [127])
may also benefit from multi-modal, distributed sensory systems; because as the physics
of the environment change, the underlying sensor modes and distributions may have to
change accordingly.
The best sensory modalities were related to known dynamics of compliant fins.
Bending sensation was more useful for force prediction than pressure sensation in the
compliant, multi degree of freedom, robotic fin (Table 7). Bending and compliance have
been shown to play a major role in force production in fins, and good estimation of the
forces occurred using distributed strain measurements. Past study of bending and
curvature in sunfish pectoral fins show that the fish can modulate fin stiffness [16, 17].
Since stiffness control is a major mechanism of force production in fins [17], it is
reasonable that bending would be highly useful for force prediction, as demonstrated
above empirically. The importance of bending measures in fins is consistent with
behavioral biology studies [3] and neurobiological evidence [6]. Theories of haptic
function have argued that local strain measures may be more informative for contact
sensing than pressure measures and may be what is more commonly found in biological
from computational biology, neurobiology, and haptics disciplines can provide insight
into sensory instrumentation in robotic systems. However, in this study, bending data
alone was not sufficient to predict propulsive forces in all robotic swimming modes
(Table 6), so propulsive force prediction may perform best using multiple sensory
modalities, such as: fin ray bending, fin pressure, membrane stretch, and further
components of the strain.
The placement of sensors is important for force prediction. Good placement may
agree with the areas responsible for force production in animal and robotic models.
Sensors along the dorsal leading edge, both bending and pressure sensors, were essential
for good prediction of propulsive forces. The importance of the dorsal leading edge in
thrust force production has been well documented in study of bony-finned fishes [11, 32].
Since sensor placement was determined primarily by study of steady swimming modes
[70], it was consistent that model performance was highest for predicting force
magnitudes during steady swimming. Further, biological evidence suggests that free
nerve afferents in sunfish pectoral fins innervate multiple regions of the fin and have
relatively high density in the dorsal and ventral leading edges [6, 19, 20]. In this way,
engineers may be able to look at simulation data and biological studies of relevant animal
gaits in order to determine initial locations for sensors on their biologically-derived
robots.
Distributed, heterogeneous sensors in fins can serve multiple roles in the control of
propulsive force: they can be used for forming accurate predictions of the force (Figure
39), can serve as direct inputs to kinematic controls, and can be exploited for more robust control with redundancies. In some cases, the sensory data could be used to predict the
propulsive force in advance (60-70ms; Figure 44), which could be useful in a
feedforward control framework. While sensory data can be fused to predict forces, the
individual sensor components are still available to the operator and can be used in a
control framework for fast updates to kinematics. Intrinsic sensors provide information
about the strain and pressure distributions across the fin’s control surface, which can aid
in understanding the components of the propulsive force; whereas this local information
about force is largely unavailable to a single extrinsic sensor (such as a force sensor at the
fin-body interface). Having access to local components of the propulsive force (local
strains, pressures) can be exploited by a controller that updates local kinematics
according to these measures. Using both the force predictions and the sensory data
directly provides more feedback pathways in a control framework. Intrinsic sensors
provide redundant information due to the mechanical coupling created by the fin
webbing. Redundant sensors can be used to help localize contact with an obstacle (as in
terrestrial examples of whisking [65, 66]), to compensate for a sensor failure by sampling
from surrounding sensors, or to provide weighted estimates of the force that improve on
estimates from single sensors or single modalities (c.f. robotics research in hyperacuity
[84]). The redundancy gained by through intrinsic, distributed, heterogeneous sensors can
be exploited for more robust estimates and more robust control. These advantages are
ripe for exploration in future work with biologically-inspired robots.