Active object perception using Bayesian classifiers and haptic exploration

Abstract

To recognise objects using only tactile sensing, humans employ various haptic exploratory procedures (EPs). Because the time, effort, and information acquisition costs of different EPs vary, choosing the best EP for accurate and efficient perception is usually based on prior knowledge or experience, also known as active exploration. An active EP selection algorithm based on a Gaussian mixture modal and Bayesian classifier has been developed to empower robots with similar intelligence. To choose the best EP for the next perception iteration, the information gain and total time cost of all actions required to identify the object are both considered. Six EPs were realised using a designed robotic arm platform, allowing eight features representing the object’s surface and geometric properties to be extracted. To evaluate the algorithm, offline data and real-world experiments were used, with the random method as a comparison. According to the results, the active method outperformed the random method with higher accuracy and in significantly less time. It had an average of weighted information gain of 132.6 and a time cost ratio (spent/total time) of only 0.3.

Appendix: Validations using the UCI data

To verify the feasibility of the proposed method, samples obtained from the UCI Machine Learning Repository (from www.archive.ics.uci.edu/ml) were also used. Iris data, which contains 3 objects, each with 4 features, and white-wine data, which contains 5 objects (wine quality), each with 11 features in total, were used for validations. For the white-wine data, 4 objects (quality 4–7) and 8 features were selected, as some features were deemed too similar and others lacked sufficient samples, as will be described in better detail below.

The initial features in the previous simulations were consistent, however, in this section, an initial feature selection logic is introduced, based on the learned data and trained classifiers, to decide which are the best initial feature(s). Due to a lack of information in this test, the time required to acquire each feature has been set as a constant. However, it is changeable based on the practical feature sample extraction procedure.

When deciding the optimal initial feature, the mean value of one random feature $f_i$ from all the features F is used. Another feature $f_j$ from one of the objects is subsequently selected to form the feature vector ${\textbf {f}}_v=(f_i, f_j)$. The weighted information gain is then calculated with the corresponding function mentioned in Sect. 2.2. Traverse all the features except $f_i$ as the new $f_j$ to find the feature identity(ID) that provides the largest WIG as the best feature. Repeating the process with all $f_i \in {\textbf {F}}$ as the initial features and recording all feature IDs with the largest WIGs. To decide the overall best initial feature, a feature number vector ${\textbf {f}}_{n}$ is used to present the number of times each $f_i \in {\textbf {F }}$ appeared as the best feature. After checking one object, the same process is repeated using another object’s features, and the ${\textbf {f}}_n$ is updated. Once the data of all the objects has been examined, the final ${\textbf {f}}_n$ is obtained, and the feature with the largest number in ${\textbf {f}}_n$ is identified as the most useful. The method is described in detail in Algorithm 3.

Using the initial feature selection method, the best initial features were determined for the iris and white-wine data. Feature 4 was chosen as the initial feature for the iris data because it provided the largest WIG 9 times. In contrast, the initial feature 8 was the best for white wine samples.

To train all possible classifiers, 40 samples for each iris feature and 125 samples for each white-wine quality feature were selected during the training phase. Forty samples were used for each feature of both objects during the testing phase (there was no validation process). Both the random and active methods were used for object recognition tests, and the results were recorded according to Sect. 3.1.

Fig. 7

The result of one object recognition process with the UCI white-wine sample. The total number of EPs is 8. The figure depicts both the confidence and error of the random and active methods. For the random method, 7 EPs were used until the confidence exceeded the threshold (0.9). and it was successful as the error dropped to 0.083. Meanwhile, the active method used the feature 8 as the initial feature, and the confidence rose from 0.425 to 0.99 after 3 EPs (3,7,8). It shows that the active method can recognise the object more efficiently

The detailed results are shown in Table 6. It can be seen that for the iris data, the recognition accuracy is 100% for both methods since there are significant differences between the features, however, the active method outperformed the random method in terms of both time cost ratio and used EPs, especially for iris Virginica which was approximately half the cost. For the white-wine data, the recognition accuracy decreases due to the similarity of the feature values, particularly quality-6, with only 55% for the active method, and 35% for the random method. This may be due to the fact that quality-6 has the most feature samples with a broad value range, and is similar to quality-2 and quality-4 features, making recognition more difficult. In total, average accuracy of the active method is 77.5%, while the accuracy of the random method is 52%, indicating that the active method performs better. Moreover, the active method has the best time cost ratio and EPs.

Figure 7 shows one recognition process result of both the random and active methods for the white-wine sample, with quality-7 serving as the ground truth. As can be seen, the results are comparable to the simulation of artificial data in Fig. 7. In addition, with the initial feature selection method performed beforehand, the active method computation time decreases significantly (Table 3), for two reasons, firstly, the initial feature selection time cost has not been added, and secondly, it helps to increase the initial confidence.