Professional Documents
Culture Documents
Sean Anderson
Sean Anderson
DSCC2018-9140
ABSTRACT data at each robot position. As the robot moves through the en-
Robotic mapping and simultaneous localization and map- vironment, a large volume of data is collected and analyzed (typ-
ping (SLAM) typically rely on sensors that produce a large num- ically on-line) to produce accurate and useful maps. If, in addi-
ber of measurements at many locations in an environment to pro- tion, the pose of the robot is unknown a priori, then the problem
duce an accurate map and, in the case of SLAM, the pose of the is one of Simultaneous Localization and Mapping (SLAM). The
robot in that map. However, with the advent of small, low-power literature on SLAM is vast and a comprehensive overview of its
robots with insect-scale features, there is a need for techniques many incarnations and variations can be found in [2].
that can produce useful maps using limited capability sensors This paradigm of acquiring large sets of data to produce
and a small number of measurements. In this work, we focus maps, however, is not applicable in every robotic scenario. In
on the use of compressive sensing to extract local environment recent years, there has been increasing interest and ongoing de-
reconstructions from ultrasonic sensor measurements. We first velopments in very small, low power robots such as autonomous
examine a simplistic setting where a square pulse is emitted and flying micro-robots (RoboBees) [3], micro air vehicles [4], bat-
use the returned echoes in a compressive sensing scheme to re- like robots [5], and insect-scale flapping wing robots [6]. It is
construct the locations of objects inside the sensing cone. We not feasible to carry or to power large sensors on such platforms
then extend this to the more practical setting, accounting for the nor to handle or store large sets of data. Even without the ability
wave nature of the acoustic signal and corresponding issues of to synthesize large numbers of sensor measurements into maps,
interference, showing that these can be accounted for in design- autonomous navigation can feasibly be done with very limited in-
ing the measurement matrix of the compressive sensing descrip- formation using techniques such as optical or echoic flow [7, 8],
tion of the problem. We demonstrate the performance of our landmarks [9], or similar approaches. However, it is conceivable
approach though several simulations. that one would like to deploy one, or more likely, many such
robots in an environment to collect data and then return to a cen-
tral hub to generate a large scale map. Because of limited storage
1 INTRODUCTION and power on-board, ideally one needs a light, low-powered sen-
Robotic mapping is an important component in the develop- sor and algorithms that can generate the most information out of
ment of autonomous robots. There are various of algorithms that each sensor measurement while also producing good maps from
allow robots to reliably use on-board sensors to map their envi- heavily undersampled data.
ronments [1]. Typically, mapping is done using laser rangefind- In this work, we focus on ultrasonic sensors. These are low-
ers, cameras, or other powerful sensors that provide a wealth of cost, light-weight sensors that have been used in both mapping
2.1 Measurement acquisition q̂ = arg min kqk1 subject to ky − Aqk2 < ε (2)
As stated above, m measurements are taken from the true
signal x through the measurement matrix Φ. Ideally, Φ exactly where ε is a user-defined parameter establishing an upper limit
This example fits into a broader class of CS approaches, (a) A single robot (b) Multiple robots/perspectives
the analysis model, which seeks sparsity in the product of the
weighting coefficients with some analysis operator, whereas the FIGURE 1: Simple illustration of two sensing scenarios. (a) A
more common synthesis model seeks sparsity directly in the co- single robot with four objects; three objects are in the sensing
efficient vector. As before, the analysis problem (3) can be mod- cone and will return echoes to the receiver. (b) Two robots look-
ified to account for noise by relaxing the constraints as in the ing at the same scene from two vantages. Two of the objects are
synthesis case in (2). seen by both, allowing the symmetry inherent in a single mea-
surement to be broken.
delays. A reconstruction from the received signal and the Φ ma- symmetric, the algorithm cannot distinguish between cells that
trix was generated using the greedy algorithm Compressive Sam- are equidistant. This shows up in the two closer objects; the re-
pling Matching Pursuit (CoSaMP) [36] implemented in Matlab. construction indicates that either of two cells is possible for each
It should be noted that in the absence of additional constraints target. There is, in fact, an entire family of equivalent solutions
in the optimization problem, the reconstruction coefficients can where the weighting of the reconstruction is spread between all
take on any real value. It is relatively straightforward to impose the symmetric possibilities in any convex combination. In gen-
additional constraints to ensure either non-negative, integer or eral however, the sensor will not be aligned with the grid and
even binary coefficients. It is also important to note that the re- thus in practice very few cells will share an exact distance with
construction considers each grid cell in its entirety; there is no another.
possibility of “partial” occupancy. However, by changing the There are various ways to resolve this symmetry issue. As
model and the corresponding Φ matrix, it should be possible to the goal is to generate a map of the entire environment, it is nat-
encode additional information in the reconstruction, such as re- ural for the robot to move through the environment, collecting
flectivity or object size relative to the cell resolution. information from multiple viewpoints as shown in Figure 1b,
In using CoSaMP, the user specifies the desired sparsity level and, possibly, to have multiple communicating robots working
(that is, the number of coefficients which may be nonzero), the together to generate a map. By combining overlapping measure-
vector of observations, and the combined A matrix. Outputs in- ments, ambiguity due to symmetry can be resolved. Note that
clude the vector of optimizing coefficients, the residual, and its under this approach, it is advantageous to have a reconstruction
norm. The residual is computed as the difference between the ob- that indicates all possible object locations rather than commit-
servations and the product of the selected coefficients with the A ting to any particular choice to avoid “false negatives”. This may
matrix. Greedy algorithms such as CoSaMP reconstruct one co- be particular beneficial when combined with a Bayesian environ-
efficient at a time, selecting the one that minimizes the residual. ment model to account for noise in the measurements.
This coefficient is then removed from the reconstruction problem When looking at a single set of measurements, one straight-
and the process repeated. forward way to break the symmetry is to use a pair of sensors
In our working example, the reconstruction found by with different orientations. This would address the problem in
CoSaMP is shown in Figure 5. In this figure, the three true targets several ways. First, it would enable the use of two simultane-
are shown as black dots and the reconstructed values are shown ously emitted pulses of different lengths (or more complex sig-
in grayscale. For consistency across this work, all reconstruc- nals in general) to propagate through the environment creating
tions use a 1m grid sizing and impose a sparsity level of 15 and a more unique signature for received signals. Second it would
non-negativity constraints. mean that the set of two sensors would now be able to discern
In this example, the sensor is aligned with the map grid and right from left based on intensity differences (note that in a three-
as a result there is a symmetry about the centerline of the sensing dimensional setting, a third sensor would be required to break the
cone. Because the outgoing pulse and receiver characteristics are up-down symmetry). One could conceivably also use time delays
(d) Received signal (e) Error with simple Φ (f) Error with milti-path Φ
FIGURE 8: Results with wave-based pulse and floor reflections. (a) True environment and (d) actual received signal. (b) Environment
reconstruction and (e) the error between the true and reconstructed signal pulse (i.e. y − Φq) if floor reflections are not accounted for.
Notice the many spurious object positions in the reconstruction and the distinct lack of coefficients which acknowledge the near and far
objects. (c) Environment reconstruction and (f) the error if floor reflections are accounted for. Notice the accuracy of the environment
reconstruction is again limited primarily by symmetry. The non-zero error reflects the difference between the grid locations (as captured
in Φ and the actual object positions.
8a is the true environment and Figure 8d is the measured signal Adding object-ground interference to the observation matrix
used in the subsequent reconstruction. Note that this is the same is simple enough, but in doing so there is still some assumption
environment as is used in Figure 7b. Figure 8b is the recon- about the relative strength of signals coming from the ground re-
struction using the single path model. The residual error signal, flections to those from the object itself. The major factors here
calculated as rerr = y − Aq, is shown in Figure 8e. Finally, Fig- are the ground and object surface and angle-dependent reflection
ure 8c is the reconstruction and Figure 8f is corresponding error properties. Clearly carpet and concrete floors would generate no-
signal using the multi-path model. Because of the strong de- ticeably different reflections and therefore different Φ matrices.
structive interference patterns that exist in the near and far object Note that this same essential description holds even in more
pulses, the simple sinusoidal basis elements of the single-path complex environments composed of many walls (like an office
measurement matrix are not suited to representing an object in space) or objects. However, the number of angles at which sound
the neighborhood of each target while also driving the residual can bounce off various surfaces before reaching the robot in-
error down. This is obvious in the case of the far object for which creases dramatically, complicating the analysis presented above.
the pulse has two large spikes but a small average amplitude, and The possibility of unknown multi-path modes of interference
the reconstruction algorithm prioritizes the residual error caused make the construction of a physically representative Φ matrix
by the larger amplitude pulse from the middle object. Though much more difficult. However, as long as multi-path or other
it may be hard to see in the multi-path reconstruction map, each phenomena can be modeled, they can be incorporated into the
reconstruction uses the maximum allowable number of nonzero observation matrix. It is also important to note that because the
coefficients. time duration of the signal is short interference is only an issue