Last weekend I ran an RSS workshop with Henrik Jacobsson, Danijel Skocaj, and Geert-Jan Kruijff. Today, I was writing up a breifing for the euCognition project and thought I’d also post that here. Soon, you’ll find all of the papers linked on the workshop website: http://www.dfki.de/cosy/www/events/InteractiveRobotLearning2008/
Interactive Robot Learning, RSS 2008
Many future applications for autonomous robots bring them into human environments as helpful assistants to untrained users in homes, offices, hospitals, and more. These applications will require robots to be able to quickly learn how to perform new tasks and skills from natural human instruction. The key here is to make it possible for the human to interact with the robot without having to read a manual.
The workshop on Interactive Robot Learning (IRL) was held at the Robotic Science and Systems 2008 conference. The discussion spanned the breadth of research questions at the intersection of Machine Learning and Human-Robot Interaction.
The workshop began with a keynote speaker, Jeff Orkin from MIT, with experience from the video game industry. Orkin gave an overview of a project in which he and his colleagues are collecting data from thousands of people playing a game called The Restaurant Game. In the game people act out a normal scene of being a waiter or a customer in a restaurant, thereby “teaching” the computer about social behavior and dialog that are common in this situation. One important thing that we learned from Orkin’s project related to IRL is that people will not always give perfect input or examples to your learning system. Therefore it is important to collect enough data and to use algorithms that let the anomalies wash out of the model.
In addition to invited keynote speakers we had 7 papers that were submitted, reviewed and selected to present at the workshop. In the morning session we had three of these speakers present. The first was Sylvain Calinon from EPFL, who spoke about a programming by demonstration framework that incorporates natural teaching mechanisms like paying attention to pointing and gaze direction of the teacher, and allowing the teacher to physically move the robot during training. Maren Bennewitz then presented a paper about recognizing gestures, like head nods, and hand gestures. Many of the gestures were quite generic and would have broad use in communicating to a robot learning system. Olaf Booij presented work on interactive mapping. In this project a robot learns a semantic map of places in a home, by clustering sensor data. It uses interaction with a human partner to simplify the task. In ambiguous situation it engages the human partner in a simple dialog, asking the name of the current location.
The afternoon session began with a keynote speaker, Jan Peters from the Max-Planck Institute of Biological Cybernetics. Peters works in robotics, nonlinear control, and machine learning. In his talk he covered a framework of motor skill learning for robots. This starts with parameterized motor primitives, used for movement generation, and then higher level tasks involve transforming these movements into motor commands. To achieve this Peters introduces an EM-based reinforcement learning algorithm, which learns smoothly without dangerous jumps in solution space. Additionally, learning smoothly in the solution space is likely to be the most understandable to a human teacher.
In the afternoon we had three more paper presentations. Two papers were in the realm of assistive robotics. Adriana Tapus presented a robot therapy system that learns and adapts on-line to personalize the therapy and maximize health benefits/outcomes. Their approach is a novel incremental learning method, positive results were found with both stroke rehabilitation patients and dementia patients. Ayanna Howard also presented an therapy application for interactive learning. Their goal is for the robot to be able to observe and evaluate a therapy patient’s exercises, assisting the job of the therapist. They presented two methods for learning to recognize therapy exercises from visual input. Jure Zabkar presented the final paper of this session about using qualitative representations in robot learning. Zabkar argued for this approach, as it results in models that are intuitive for non-expert human’s to inspect and understand, which is a key component of interactive learning.
The workshop ended with a final keynote speaker, Aude Billard from EPFL. Billard has made several contributions over the years, in the realm of robot programming by demonstration, and her talk covered many aspects of the work that her lab has tackled on the problem of imitation learning for robots. In particular their focus on the complementary problems of “what to imitate”, and “how to imitate.” The first is about determining the key components or features that really represent the goal of a task). Having a framework for determining “what to imitate” give the system means to generalize appropriately. Their approach is based on Gaussian Mixture Models. The “how to imitate” problem involves translating motions seen by a human to motions the robot itself can do, and also achieving the goal of the task. Their approach is a stable dynamical system, active in a hybrid cartesian-joint angle frame of reference. This ends up being able to handle perturbations in the environment, joint angle limits, and adapts to changes in target position.
July 3rd, 2008
Posted by
A.L.T. |
Machine Learning, Conferences, HRI |
no comments
What is anthropomorphism?
I think a lot of times when roboticists talk about anthropomorphism, it has an air of cynicism. Like “oh those silly people, they think the robot has thoughts and feelings.” Which is obviously not true for most people, but it’s interesting to think about why people have the propensity to anthropomorphize, and to understand what, if anything, we should do to design for this.
Dan Dennett’s Intenional Stance offers some insight. It explains that while people realize that some thing may not actually have desires and intentions, it ends up being an efficient and successful way to reason about a lot of phenomena that we come across. Put another way, as social animals, humans became very good at reasoning about mental states (beliefs, desires, intentions) to predict the actions of other animals. Human’s applied this reasoning strategy to anything that produced some self-motivated action that could not be described by physics. These days it’s not just animals, there are a lot more things in the world that produce seemingly self-motivated actions that are not (on the surface) describable by physics (cars, computers, robots, …); therefore, the mental state reasoning strategy kicks in.
Don Norman has a similar stance and has long studied how an object’s design and appearance communicate and inform people of the object’s possible functions. In a recent book he takes this into the realm of robots, arguing that anthropomorphism is an abstraction that will help people understand how to interact with a robot, and that robots should use familiar mechanisms like emotive expressions to communicate internal state to a human user. 
Maximizing anthropomorphism
The field of animation has a long history of maximizing anthropomorphism. Frank Thomas and Ollie Johnston wrote a beautiful book in 1981, The Illusion of Life, about the principles and process employed in some of the Disney classic animations. In a SIGGRAPH paper in the late 80s, Pixar’s John Lasseter describes how the Disney principles of creating the “Illusion of Life” in 2D animation should translate to 3D animation. In the paper he steps through the 11 fundamental principles of traditional animation and discusses their 3D corralate, arguing that many of the principles transcend the particular medium. Which to me begs the question….How can these principles of animation inform how we create of an “Illusion of Life” in a robot?
Animation Principles for Robots
I’ll mention six of the principles (interpreted pretty loosely) that I find most directly applicable and interesting for thinking about robot behavior design. Most of these can be summed up as: It’s not just what you do, but how you do it. In robotics, I think we tend to work on how to select a particular motor control program, action, or behavior (selecting what to do). These principles of animation argue for having additional mechanisms to dynamically control parameters of how you do it.
Appeal — Creating a design or an action that the audience enjoys watching. This is what usually comes to mind when people talk about robots and anthropomorphism. It is the principle that would argue for creating robots with baby or pet-like proportions, so they are inherently appealing and nice to watch.
Anticipation — The preparation for an action. This is the technique of doing something to prepare the audience for the action that is about to come. Directing their attention to the right part of the screen. I think the best example of this is the “wheely feet” that cartoon characters get before taking off running. Or making sure that an arm movement winds up before exerting some energy. These are definitely cues that robots of all morphologies could use to help a human partner understand more about what they are “about” to do. It would make people feel more comfortable if they felt they could accurately predict the robots behavior. Thus, I think cues to help with audience/user anticipation are important. I’m not currently working with any mobile robots, but if I was…I would be making them some “wheely-feet”!
Arcs — The visual path of action for natural movement. Many of these principles argue against motion control that is simply based on an efficient path, or some inverse kinematics. Arcs, refers to the fact that even in a reaching action or something where the end point is the goal, there is something about the path that is taken to the end point that can make the motion more natural (and, I would hypothesize more predictable. This could ease interaction because a person would find the robot’s actions more intuitive to follow).
Secondary Action — The action of an object resulting from another action. So, the part about objects moving in relation to a character’s actions, robots get for free with physics. But a more subtle aspect of secondary action is within the character’s body. As one example, when you nod your head, it is more than just your neck bone moving. A natural head nod will have secondary actions in some of the robot’s facial features (e.g., if it has eyebrows, or ears). These secondary actions, while not instrumental to the getting from A to B goal, create a more natural looking behavior.
Timing — Spacing actions to define the weight and size of objects and the personality of characters. Robots can certainly use this one. Slow motions communicate something different than fast motions. Thus, not just getting the hand from A to B, but again, it’s about how you get there with a particular speed.
Follow Through and Overlapping Action — The termination of an action and establishing its relationship to the next action. I think this relates most to the prior point about anticipation. Not only does a single action need to have “set up,” but all of the robot’s actions have to make sense together. One is the anticipation of the other. And importantly, a particular action or behavior might terminate differently depending on what action or behavior is coming next.
June 17th, 2008
Posted by
A.L.T. |
HRI |
no comments
I attended the NEW HRI workshop at ICRA. Though it wasn’t ever clear what they meant by “NEW,” it was a nicely put together workshop that generated a good discussion about “What is HRI anyway?”
All of the speakers were asked to focus on the same three questions:
1) What characteristics unify and should be common to all areas of HRI research?
2) What properties typify novel and significant research in HRI?
3) How can we further synergistic research between HRI and the broader robotics community?
I think it was good to have this workshop at ICRA because there were a broad range of people there. There were plenty of “big-H little-R I” folks, and a fair number of “little-H big-R I” folks as well. There was a *very* wide range of answers to the three questions. I think there were really only one or two people that had any overlap in what they thought were unifying characteristics of HRI research, or what should be common to all HRI research.
A lot of the discussion boils down to what should count as a contribution in the HRI community. Do you have to do a user study for it to be HRI? How interesting/complicated does your robot need to be for it to be HRI? etc.
In the end everyone agreed that it is far too young of a field to start drawing boundaries. At this point, to make sure that the field doesn’t end up being some very narrow aspect of HRI, we need to be inclusive and appreciate a variety of different kinds of research contributions.
Thus, the day ended on a very Kumbiya note. But several people pointed out that what really matters is when the rubber hits the road. For this multidisciplinary field to succeed, we need people to appreciate and advocate for a broad view of HRI contributions in conference, journal, and tenure reviews.
June 10th, 2008
Posted by
A.L.T. |
Conferences, HRI |
no comments
Well, I fell off the wagon last semester, but it’s time to dust off the blog and get things going again.
It’s been a fun, exciting, and hectic first year at Georgia Tech. I taught a new grad course in the Spring Designing the course was interesting in and of itself, a seminar on Human-Robot Interaction. Given that the field of HRI is still defining itself, the topic list for a course on HRI is pretty up in the air. So, the course reflects my particular interests in HRI and has an AI and Cognitive Science bent. I chose a few broad categories of social intelligence and for each category we did a survey of readings related to what we know about how this capabilities arises in humans, and then coupling that with readings on state-of-the-art approaches to implementing these kinds of capabilities on robots. For example: understanding intentions, social learning, teamwork, empathy and emotions.
My other main focus this year has been getting together a social robot platform for my research group. I’m working with a company, Meka Robotics, who are building us an upper torso robot with two arms and three finger hands. And a group at the Georgia Tech Research Institute is building us a socially expressive head. It’s been a really fun and interesting process, stay tuned for more details as we start to get the hardware up and running.
June 6th, 2008
Posted by
A.L.T. |
Announcements |
no comments
Jan Peters and Marc Toussaint are running a workshop at NIPS this year called “Robotics Challenges for Machine Learning.” Abstracts due Oct 21.
Creating autonomous robots that can assist humans in situations of daily life is a great challenge for machine learning. While this aim has been a long standing vision of robotics, artificial intelligence, and the cognitive sciences, we have yet to achieve the first step of creating robots that can accomplish a multitude of different tasks, triggered by environmental context or higher level instruction.
It would be great to get the NIPS community thinking about robotics, and in particular the kind of adaptation and learning that needs to happen in order for robots to be useful assistants to humans in everyday life. So, I encourage people to participate. The stated goal of the workshop is to bring together people that are interested in robotics as a source and inspiration for new Machine Learning challenges. My take is that social interaction is a new challange, and that it’s a necessary and useful constraint for Machine Learning algorithms in these scenarios. Should be a great venue for that conversation.
October 11th, 2007
Posted by
A.L.T. |
Conferences, Announcements |
no comments
Beki Grinter and other colleagues at Georgia Tech recently had a paper at Ubicom on their ethnographic study of people’s Roomba usage. It colllected quite a bit of press, an AP article, robots.net, and even inspired Comedy Central’s Colbert Report to feature robots as the #1 Threat Down…nice!
These Roomba studies always leave me wondering about the people that we are not hearing from. I personally know a lot of people that stopped using a Roomba, they either didn’t find that it cleaned very well, or got tired of “roomba-izing” their houses. It’s facinating to learn about the Roomba fans, but as a counterpoint it would be great to see some interviews and analysis of the non-fans too.
October 11th, 2007
Posted by
A.L.T. |
In the News, Fun, HRI |
4 comments
Last week Randy Pausch gave a touching and inspiring lecture at CMU. I’d like to join in the chorus of the media and blog-o-sphere and encourage people to watch it.
The short story: Randy has been fighting cancer, and recently found out that the treatment hasn’t worked, and he has only a few months. Randy is famous for co-founding the Entertainment Technology Center at CMU, and for directing the Alice software project. The lecture is a very thoughtful reflection on his life and work and is an inspiration.
September 27th, 2007
Posted by
A.L.T. |
Uncategorized |
3 comments
Some colleagues from MIT, Cory Kidd and Aaron Edsinger, had live demos on Good Morning America this morning. Very impressive. Robots working outside of the lab, and interacting with hosts of a TV show!
They highlighted a couple of the big social robots themes, healthcare and eldercare. A highlight was definitely when Diane Sawyer was moving Domo’s arm around asking if it could “feel” it (compliant and force sensitive manipulators was Aaron’s focus with Domo) …and when she pulled too hard Domo said “ouch.”
September 27th, 2007
Posted by
A.L.T. |
In the News, Fun |
no comments
I recently got back from the RO-MAN conference in Korea, here’s the hightlights.
Language Learning Keynote — Luc Steels gave a nice talk about language learning. His goal is for machines to be able to really speak and understand language, and his approach is that of emergent communication. The idea is that language is never perfect, people communicate by continually repairing misunderstandings and continually building up a common ground or common language, “inventing” new terms, etc. This is much different than a corpus based machine learning approach that assumes a stationary environment and other such things. So, his platform for studying this kind of emergent language development, or learning all the time approach for language, is language games for robots. He has some interesting examples with QRIO robots. It was a bit unclear how the all important repair phase was instigated or took place. But it’s great to see such an interactive alternative to natural language processing.
Human-Robot Interactive Teaching — I was in a session organized by Joe Sanders and Chystopher Nehaniv. You can read more about my talk. I enjoyed Sylvain Calinon’s talk in the session about “Active Teaching”. There are a lot of people working on robot “Learning by Imitation” But his work addresses an important question about the nature of an imitation interaction. Human imitation learning is an active process. Someone is teaching you something, they demonstrate it, you try it, you get it wrong, they go into more detail on a particular aspect, and on and on. Sylvain has some nice work on teaching a robot a task by showing it the physical movement. It then copies the movement, and you can stop the robot and select particular motors, and physically move it through the motion during the correction process. There’s more work to be done on the interaction, and I don’t think they’ve tried it with non-expert users yet. But this is work to keep an eye on.
The Uncanny Valley — I’m sorry, but I missed the uncanny valley boat. I understand the qualitative assessment that there is a class of things (like zombies, and almost human things) that people find creepy. And that many robots fall into this category either by the way they look or the way they behave. But from what I have seen and read, it seems that we all basically agree that 2 dimensions are really not enough to represent what it is that people find creepy. And that this 2-d graph of Mori’s is not based on quantitative data that we can measure our robots against. So, I’m always amazed at how many people show graphs of the uncanny valley in their talks. Someone, please enlighten me in the comments, why should I care about this uncanny valley graph?
Robot design workshop — This session was one of the better ones that I saw. Prof. Myung Suk Kim has a lab at KAIST that does some great work on robot design. He gave a presentation giving an overview of the work going on in his lab. Jodi Forlizzi talked about 5 aspects of designing for HRI: gaze (social attention), speech/sound, gesture, motion, and personality. And Tomotaka Takahashi, a PhD student at Kyoto university, gave a talk about two robots that he has built, the Chroino and the FT (for female type…). They are pretty cute, and the Chroino was licensed and developed into a product: Manoi.
AUR won — Guy’s robot, the AUR robotic lamp, won the robot design competition!
Four Keepons dancing! — Hideki Kozima and Marek Michalowski had a nice showing in the demonstrations event, four keepon robots side-by-side dancing away and interacting with people. They were in the middle of a Keepon tour, first winning the Robots at Play design competition in Denmark, then RO-MAN, and now they’re at NextFest in LA with Wired. Wired likes Keepon so much they did a response video with Spoon. It’s good clean robot fun.
Meeting a reader — I met one of the readers of this blog, which doesn’t happen in person very often, so that was fun! She was concerned that I’m going to quit blogging once I get busy with the new job. But, that’s not the intention, I plan to make time to keep the Social Robot research conversation going in the blog-o-sphere.
I really liked Korea, the people were extremely friendly. I will definitely have to make a trip back, perhaps to visit Robot Land!
September 11th, 2007
Posted by
A.L.T. |
Uncategorized |
no comments
Class is in session….well I’m not teaching until next semester, but I have officially started my job at Georgia Tech. I think that makes me a Ramblin’ Wreck, but I’m still figuring it all out.
You can visit my new digital home, which has my new contact info and not much more… hopefully it will get an update soon!
August 20th, 2007
Posted by
A.L.T. |
Announcements |
no comments
