Lessons from a robotics entrepreneur

Due to a variety of activities at work, I have been following the activities of some of the alumni of Willow Garage. In part, I am curious about what they do in terms of products and technologies, in light of what they learnt from building and popularising robots like the PR2. 

Recently, I came across this post written by Steve Cousins, who is now the head of Savioke. In this post, he takes on the question about lessons learnt:

http://www.savioke.com/blog/2014/7/23/top-ten-things-i-learned-at-willow-garage

It is clearly written from the industry perspective, but I think it is highly relevant to academics too. After all, a lot of the recent attention we are getting as a field is not so much because we have created something new that simply did not exist 20 years back. Instead, it is because even some of these new technologies in our field are finally looking close to being commercially relevant.

Winograd schemas

Through some conversations at AAAI last week, I came to know about Winograd Schemas and a challenge involving their interpretation that is considered by many to be a more fitting one than the conversation bot versions of the Turing Test that have repeatedly been shown to be easy to crack without really solving the main problem. A Winograd schema is based on pairs of sentences that differ in only one or two words which make a radical difference in the meaning, thus requiring deeper understanding of ‘commonsense’ in order to resolve the meaning.

This challenge has now been picked up by Nuance, who see substantial benefits from this level of AI to their own products. This is a positive thing.

I wonder if we can also define equally crisp and well defined challenges that bring out the need for intelligence in the area of robotics. The nice thing about the Nuance Winograd challenge is that it gets at the heart of the issue in a bite-sized fashion, allowing anyone with a keen mind and some bandwidth to potentially contribute to the issue. In contrast, most robotics challenges out there today are as much a test of people’s political and management skills in assembling the resources to compete as it is a test of scientific ideas – and even then, they often seem to be set up to not really address these kinds of core AI issues. It’d be a lot of fun to find more bite-sized yet deep and significant open problems within robotics.

Intention prediction among goalkeepers

How is it that goal keepers manage to ever save any penalty shots, beating the striker in the incredibly little time available?

This brief video in the online version of The Economist outlines a few quite different attributes to their thought process, not to mention the not so explicitly conscious reflexes. Even at this cursory level, it is interesting how many different modalities are involved – learning from the striker’s historical kicks, randomised strategies in the spirit of game theory, face processing to extract subtle cues, psychological intimidation, etc.

My student, Aris Valtazanos, and I wondered about this problem in one of our papers associated with our football playing robots, but clearly we are unable to capture this whole variety of interactive intelligence. It would be cool when one day we have agents that can actually function at this level!

 

Chess Metaphors

Chess Metaphors

This is a nice book review written by Garry Kasparov, insightful in part because of the major role he played in the area. Some interesting snippets below.

On human-computer team play:

The chess machine Hydra, which is a chess-specific supercomputer like Deep Blue, was no match for a strong human player using a relatively weak laptop. Human strategic guidance combined with the tactical acuity of a computer was overwhelming.

The surprise came at the conclusion of the event. The winner was revealed to be not a grandmaster with a state-of-the-art PC but a pair of amateur American chess players using three computers at the same time. Their skill at manipulating and “coaching” their computers to look very deeply into positions effectively counteracted the superior chess understanding of their grandmaster opponents and the greater computational power of other participants. Weak human + machine + better process was superior to a strong computer alone and, more remarkably, superior to a strong human + machine + inferior process.

On the methodology of chess agents:

Like so much else in our technology-rich and innovation-poor modern world, chess computing has fallen prey to incrementalism and the demands of the market. Brute-force programs play the best chess, so why bother with anything else? Why waste time and money experimenting with new and innovative ideas when we already know what works? Such thinking should horrify anyone worthy of the name of scientist, but it seems, tragically, to be the norm.

 

Eugene Goostman and the Turing Test

I just heard about this program that participated in one of the conversational Turing Test competitions, through this article in The New Yorker by Gary Marcus.

Nobody who seriously works on any aspect of AI would be genuinely surprised by this. However, it is a useful reminder of exactly ‘how much’ AI one needs in any practical application. There remain many hard problems, e.g., natural language understanding at a human-competitive level. However, there are many applications where the bar is actually really low, e.g., as Gary Marcus notes in the above article,

If Goostman can fool a third of its judges, the creation of convincing computer-based characters in interactive games—the next generation of Choose Your Own Adventure storytelling—may be a lot easier than anyone realized.

Over the years, I have been surprised, and a tad disappointed, at how many applications that could be so stimulating for AI research are actually cracked by simplistic and naive methods, with just a bit of clever wrapping (like the game AI comment above). I wonder if there are any genuinely intermediate level problems – something not as trivially solvable by naive methods (e.g., the game AI in many apps today), yet not as steep as ‘full’ NLU?