IEEE Robotics & Automation Magazine - June 2011 - 92

they may be both found to indicate the presence of a user in a
room. Thus, higher-level reasoning may remain identical if
these two different concepts are first matched.
Van Kasteren et al. extended transfer-learning methods
to operate on time series resulting from the activation of
simple binary sensors [59]. They applied this method to
transfer behavior-recognition capabilities from one smart
home kind to another with different and a priori unknown
number and placement of sensors. The system first automatically finds how sensor activations in different environments relate to identical higher-level concepts using
statistical approaches. A recognition system can also learn
internal hierarchical representation of activities or concepts [60], upon which reasoning is performed. Hu et al.
further report on using Web mining to match concepts
[61]. Advances in merging concepts in ontologies [62],
[63] support the transfer of activity-recognition reasoning
across different conceptual spaces.
In robotics, these principles may allow the robots to
exchange the knowledge they have individually gained
about the world. This may be especially relevant when
principles of autonomous mental development are used, as
robots can develop distinct world representations according to their capabilities.
Other Approaches
Some approaches do not fit in the taxonomy above. Blanke
and Schiele proposed a form of transfer learning to reuse
action primitives across
* different but related application domains. ActionSemisupervised learning
primitive spotting (hamallows to combine a limited mering, screwing, and cutting) was trained on the
data set of a shelf-assembly
number of labeled data
task. These primitives were
reused as is to detect higherwith a large amount of
level steps of a mirrorassembly task, thus considunlabeled data to train
erably reducing the amount
of training data needed for
classifiers.
the new task [32].
*
Most of the approaches
previously described attempt to reduce or eliminate the need for training data for
activity recognition on a new platform. Beigl and coworkers
proposed to "crowd-source" the acquisition of training data.
They addressed the issues related to shared data labeling by
developing a framework suitable for end users operating on a
mobile phone [64]. Semisupervised learning allows to combine a limited number of labeled data with a large amount of
unlabeled data to train classifiers. It was successfully used to
train activity-recognition systems using only sparse activity
labels [65]. Recent trends seek further reduction in the datacollection efforts by automatically generating activity-recognition models from online sources by data mining [66].
92

IEEE ROBOTICS & AUTOMATION MAGAZINE

JUNE 2011

Calatroni et al. argue that many existing sensors can be
repurposed for activity recognition, although they were
initially deployed for other uses [67]. They show, for
instance, how reed switches placed in windows for security
purposes can be used to infer standing or walking by
means of assumptions about human behavior when interacting with the instrumented object. They indicate several
other sensors and behavioral assumptions that allow to
obtain sporadic labels about the modes of locomotion of
the user or his or her gestures. They suggest to incrementally train the body-worn recognition system whenever
such labels are obtained, with the transfer-learning method
described earlier. Eventually, the wearable system becomes
capable of activity recognition even when the user does not
interact with the source of labels. Since this process can be
continuous, the system can perform activity recognition
with many unforeseen combinations of on-body sensors as
long as they provide discriminative signals.
Finally, in wearable computing, the user and the technical system are tightly interacting. Thus, in some cases, the
user may provide information to the wearable system, such
as whether it correctly identified the last activity. This can
be used in an online learning paradigm to refine activity
models according to the user's expectations. We showed in
[68], how a binary true/false feedback can be used to guide
adaptation. In the example, the feedback was provided by
an electroencephalography (EEG)-based brain-computer
interface, but a simple push-button feedback is also possible. In robot teams, this may be akin to robot sporadically
judging the performance of their peers when performing a
collective task.
Conclusions and Outlook
Activity recognition enables a WWW for robots by providing a tool to label large robot-generated activity data sets,
by enabling activity-aware HRI, and by opening the way to
self-learning autonomous robots capable of monitoring
their own proficiency at a task. The large data sets of collective human behaviors collected in wearable computing
may also be used to bootstrap humanoid robot behaviors
and thus achieve more humanlike interactions in human-
robot societies.
Human activity recognition has been a major object of
research in wearable computing since the mid-1990s. We
summarized the methods developed in the community
along the ARC, which is a set of processing principles followed in most activity-recognition research. Since human
activities are highly variable, we reported some of the
recent advances to enhance the robustness of activityrecognition systems when they are shared among different users or deployed in different application domains.
Human activity recognition from on-body sensors is far
from a solved problem. Some of the continuing challenges include:
l finding more efficient sensor modalities for activity
recognition. They should satisfy multiple requirements:

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - June 2011