This year’s O’Reilly AI conference in San Jose, CA delivered the usual mix of exciting presentations and dynamic discussions, as well as a healthy dose of industry networking. Appen is proud to have been given the opportunity to share some recent research at the conference that promises to improve the way robots interact with people.
In their presentation Creating Autonomy for Social Robots, Dylan Glas, Senior Robotics Software Architect at Futurewei Technologies, and Figure Eight Machine Learning Scientist Phoebe Liu addressed the challenges faced in developing robots that have human-level depth and nuance in their social interactions.
Goals for creating more social robots
The researchers focused their efforts with a few key goals in mind for developing more socially intelligent robots:
- Interact with people as peers rather than devices
- Communicate using both speech and gesture
- Provide useful services that a would normally require a human being
Developing improved AI-based technologies for training social robots could mean giving them the ability to operate independently in situations similar to those faced by interacting with people and environments with a high degree of mutability. In other words, a social robot should be able to engage in complex conversations where topics evolve and change unexpectedly and social rules are implicit rather than clearly defined.
How do we do this?
Glas and Liu set out to identify ways to train machine learning models based on real-world scenarios, full of ambiguity and variety. By focusing their research in this way, they hoped to:1. Learn about people: Learn how people behave to gain the ability to anticipate and respond to them2. Learn from people: Learn social behaviors from people’s explicit and implicit knowledge3. Learn withpeople: Improve and personalize interaction logic through real interactions
Let’s go to the mall
To learn more about people, the team decided to use interactions over two years at a shopping mall in Japan. The goal of this experiment is to train a machine learning model to approach people and offer directions or recommendations for shops to visit. Glas and Liu identified several potential obstacles:
- Robots need to move slowly for safety purposes and so may not approach people with the speed required to capture their attention.
- Sensing is short-range so robots would need to be close enough to a shopper to gather required data about them to make appropriate decisions.
- Many people are simply too busy or uninterested to interact with a robot.
These concerns were relatively minor, though, compared to the fact that they would need to train a machine learning model to accurately anticipate and choose people to approach. Glas and Liu decided that they could crowdsource pedestrian data – set up data collection in a place where lots of shoppers were walking through, then used a combination of supervised and unsupervised machine learning techniques to analyze that data to build a machine learning model. Data collected around walking styles (casual, wandering, fast/ busy) and trajectories could be analyzed to identify how people’s behavior varies in different areas of the shopping center at different times and on different days (i.e. weekday vs. weekend). The goal of this being that a robot could anticipate pedestrian behavior and inform decision making around who to approach.
In addition to trajectory data, Glas and Liu suggested a flowchart-based visual programming language with blocks representing back-end functions like talk, shake hands, etc. could be used the robot’s interactions once a person had been approached and engaged. Using this strategy, they hypothesized that content authors could easily combine speech, gesture, and emotion to manage dialogue and build engaging interactions.
Crowdsourcing human behavior for data-driven human-robot interactions (HRI)
During their presentation, Liu and Glas further discussed the benefits and challenges of crowdsourcing the work of training better social robots. Annotating social behavior can be difficult because it is often subjective and fuzzy. In other words, if we can’t even clearly articulate the reason for why we adhere to certain social rules. It is also challenging to communicate instructions to workers to accurately annotate the data.
The camera shop experiment
To explore how they might address these crowdsourcing challenges, Glas and Liu set up a simulation in a camera shop where participants role played as either a customer or a shopkeeper and spoke and acted naturally without any script. Key aspects of the experiment included:
- Multimodal interaction with speech, locomotion, and proxemics formation
- A human position tracking system, based on RGB-D depth sensors, reports people’s movement in terms of X, Y every second
- Role-players also carried an android phone for speech recognition to compare performance
One thing the experiment showed is that natural conversation is very noisy; even the state-of-the-art speech recognition system reported only correct match or very minor errors in 53% of the interactions. A large amount of speech data contained some minor errors or major errors, or was totally unrecognizable. Another key finding was that there are natural variations that arise from person to person, even though they all represent the same action. For example, people use different phrases to say things that are semantically similar. But even though there are natural human variations in behavior, at its core, there are some commonly repeatable patterns people usually perform, meaning it’s possible to abstract data into representations of typical behavior elements.By introducing subtle personality changes in the role-playing experiment, Glas and Liu were able to go deeper into the subject for more nuanced data. For example, a shopkeeper was simulated using training examples with two people: one who had a quiet, shy personality, and another who was quite outgoing. These role players were asked to act as a passive and proactive shopkeeper respectively to see if it was possible to capture their interaction styles from the training data.
Learning from human behavior is critical for developing autonomous robots for the real world
Glas and Liu wrapped up their presentation by highlighting some key takeaways from their research:
- Understand behavior patterns so we can react to them and anticipate what people will do in the future so we can plan.
- Use people’s explicit knowledge when it is available and capture and model people’s implicit knowledge when it is not.
- Offline learning is only a starting point and there is a need for personalizing and adapting to social behavior in real time.
What’s next?
Glas and Liu’s work in this area is a great example of how rapidly AI technologies are advancing. We’ve reached a point where the interactive capabilities and usefulness of social robots could play important roles in diverse fields. From medical science to manufacturing, AI-powered robots are no longer the stuff of cartoons or science fiction; they look to be increasingly ready to transform the fabric of our personal and professional lives.—At Appen, we’ve helped leaders in machine learning and AI scale their programs from proof of concept to production. Contact us to learn more.
