0 5 10 15 20 0
20 40 60 80 100
Slope (deg)
Slip (%)
Unsupervised: Err=15.8 %
0 5 10 15 20
0 20 40 60 80 100
Slope (deg) Autom. sup.: Err=9.99 %
0 5 10 15 20
0 20 40 60 80 100
Slope (deg) Human sup.: Err=9.91 %
Soil Asph Gravel
Figure 3.11: Field data test results for one of the runs. The learned nonlinear slip models superimposed on the test data when learning without supervision (i.e., unsu- pervised) (left), when learning with automatic supervision (middle), and when human labeling is used (right). The test errors for this run are given atop each plot.
with automatic supervision outperforms the purely unsupervised learning and closes the gap to the learning with human supervision. More precisely, learning with auto- matic supervision achieves about 70% of the possible margin for improvement.
The learned nonlinear models and the corresponding test errors for the three terrain classes for one of the runs are given in Figure 3.11. We can see that the unsupervised learning could not learn the correct models well because of classification errors in the vision space. One should also note the large slip error even when training on manually labeled terrain types. This is because the field data is very noisy.
In this experiment we see again that learning with automatic supervision outper- forms the unsupervised learning and is close to learning with human supervision. In summary, learning with automatic supervision has the potential to substitute the ex- pensive, tedious, and inefficient human labeling in applications related to autonomous navigation.
The development of this framework is motivated by the problem of autonomous nav- igation without human supervision. An important outcome of the algorithm is that the expected mechanical behavior can be predicted from only visual or other onboard sensors and that the learning is done completely automatically.
We have shown experiments on a dataset, collected while driving in the field, in which different terrain types are learned better from both vision and slip behavior supervision, than with vision alone. The impact of the proposed method is that it can enable the rover to drive safely on slopes, learning autonomously about different terrains and its mobility limitations on them.
3.6.1 Limitations and future work
In this chapter, we have limited the test scenario to three terrains, because the re- maining two terrains had a range of missing values. Extending the work to be able to handle missing values in the data, as shown in Figure 3.12, would be important in expanding the applicability of the approach in practice. As seen, for the sand terrain there are no measurements of slip apart from ∼0 degree slopes. Since the observed slip for these slopes is already quite large, the missing values indicate that the vehicle cannot traverse larger slopes. This information can be exploited during learning as well.
In this work we have assumed a simpler form of the slip models so that they can be learned with a reasonable amount of data. Extension to supervision which can consider the other variables slip depends on (e.g., both slopes, roughness, etc.) is also possible. However, a much larger amount of data for learning is needed. Extension to other forms of supervision, e.g., vibration signals [28, 37], etc., would be also useful from a practical standpoint.
An important future direction is extending the approach to online learning. This problem is quite challenging namely because of the ambiguity of the slip measure- ments. For example, if the same slip has been measured this does not necessarily mean that the terrain type is the same. The main challenge in this case is deciding
−10 −5 0 5 10 15 20 0
20 40 60 80
100 Slip space data
Slope (deg)
Slip (%)
SandSoil Gravel Asphalt
Figure 3.12: Missing data problem extension. Slip measurements plotted as a function of the estimated slope angles retrieved from actual rover traversals. Data for some slope ranges may be unavailable, as is true here for the sand terrain. This may be indicative of the rover not being able to traverse any larger slopes and may be sufficient information to learn the whole slip model. LAGR vehicle
on the appropriate data to store in memory and making a decision when to create or merge two slip models.
Extending the work to high-dimensional input spaces is an important direction of future work since more complex feature spaces are needed to better describe natural terrains. We develop such an algorithm in Chapter 4.
Chapter 4
Dimensionality Reduction from Automatic Supervision
In this chapter we address the problem of nonlinear dimensionality reduction in which some additional signals, possibly noisy and ambiguous, are used as supervision. Sim- ilarly to Chapter 3, we address the problem of learning to recognize different terrains in a fully automatic fashion, using the robot’s slip signals as supervision. In this chapter, however, we consider much more complex representations which are more descriptive, but are also technically more challenging since they are high dimensional.
We present a novel probabilistic framework, based on Mixture of Factor Analyzers (MFA) [43], in which the additional supervision affects the construction of the lower- dimensional representation of the input data [5]. Intuitively, incorporating supervision into the dimensionality reduction process can help create a lower-dimensional repre- sentation that better separates clusters which are close in the original space, but need to be discriminated for the task at hand. From a practical standpoint, incorporating supervision into the dimensionality reduction process is important, as some terrains might be visually similar but induce very different robot mobility, and choosing a lower-dimensional visual representation with respect to the supervision will improve the vision-based terrain learning and the final classification performance.
The proposed approach is used for automatic vision-based learning of terrains for the purposes of slip prediction. It has been tested on field test data collected by the LAGR robot while driving on soil, gravel, and asphalt. Our experiments show that
using additional supervision in the dimensionality reduction process contributes to improving the classification and slip prediction performance over the unsupervised dimensionality reduction.
This is the first work that proposes supervisednonlinear dimensionality reduction in a probabilistic framework usingautomatic, noisy and ambiguous, supervision com- ing from the robot’s sensors. The proposed method stands in between methods for reasoning under uncertainty using probabilistic models and methods for learning the underlying structure of the data.
4.1 Introduction
As in the previous chapter, we address the problem of learning to recognize terrain types when the rover’s slip on different terrains acts as supervision for learning. In this chapter we consider a more realistic high-dimensional representation of the input vision data. This presents a novel challenge, as working with high-dimensional data requires a prohibitive amount of training examples, knowledge about the underlying structure, and might incur numerical computation issues.
Complex high-dimensional representations are common in applications related to vision, document retrieval, and robotics [47, 48, 102]. Although such data are of high dimensionality, they have been shown to reside on lower-dimensional manifolds [16, 102, 115]. Projecting the initial data to a lower-dimensional space can alleviate the learning task, e.g., by bringing related data points closer together or by requiring fewer training examples. Numerous methods for nonlinear dimensionality reduction have been proposed: Isomap [115], Locally Linear Embedding (LLE) [102], Laplacian Eigenmaps [16], Locality Preserving Projections [52], etc.
Nonlinear dimensionality reduction techniques are generally unsupervised [16, 43, 102, 115], as they have been intended mostly for data representation. In this chapter we propose to learn a more useful lower-dimensional visual representation which at the same time allows for better discrimination of terrains, determined to be different by the automatic mechanical supervision from the robot. The intuition is that, in
practice, some additional information, regarding which data points are more similar and should cluster together, might be available and could be exploited to obtain better low-dimensional representations. For example, the autonomous robot can provide additional information about traversability on different terrains from its mechanical sensors which can complement the already available visual information about the terrain. Such signals could be potentially very useful to discriminate terrains which are visually similar and could not be discriminated from vision alone.
In this work we propose to do supervised nonlinear dimensionality reduction in which noisy or uncertain signals act as supervision. We present a probabilistic frame- work in which the additional supervision can influence the selection of more appro- priate and meaningful low-dimensional projections with respect to the learning task.
We incorporate the raw sensor signal measurements into a framework for reasoning under uncertainty, in which the most likely decision is made, taking into consideration all the observed data. This allows working with the supervision signal to be noisy or ambiguous, i.e., it might not have a one-to-one correspondence to the input data.
An important idea of the approach, from a learning perspective, is that the lower dimensional representation can be selected in a more appropriate way regarding future classification tasks. The significance of the approach, from a practical standpoint, is that fully automatic learning and recognition of terrain types can be performed without using human supervision for data labeling and that the method can work with high-dimensional representations Moreover, the method allows the supervision signal obtained by the robot to be noisy or ambiguous.