Humanoid Robots For the Next DARPA Grand Challenge?

HizookRobotics writes "The official announcement should be out very soon, but for now here's the unofficial, preliminary details based on notes from Dr. Gill Pratt's talk at DTRA Industry Day: The new Grand Challenge is for a humanoid robot (with a bias toward bipedal designs) that can be used in rough terrain and for industrial disasters. The robot will be required to maneuver into and drive an open-frame vehicle (eg. tractor), proceed to a building and dismount, ingress through a locked door using a key, traverse a 100 meter rubble-strewn hallway, climb a ladder, locate a leaking pipe and seal it by closing off a nearby valve, and then replace a faulty pump to resume normal operations — all semi-autonomously with just 'supervisory teleoperation.' It looks like there will be six hardware teams to develop new robots, and twelve software teams using a common platform."

Re:Bipedal humanoid

[HizookRobotics]

This is a common misconception. In academia, "humanoid" will usually refer to a robot with _some_ humanoid features (eg. two arms). If you look at the "Humanoids" conference, you'll see that there are a healthy mix of legged and wheeled designs. So I thought specificity was good in this case...

That's ambitious

[Animats]

This is very ambitious. Yet it's almost within reach, if enough money is thrown at the problem. A lot of money.

First, this requires solving low-speed legged locomotion over difficult terrain. Walking over rubble and climbing a ladder will be very tough. But it's not totally out of reach. A machine which can walk on either two or four limbs would be an advantage over rubble. Ladder climbing is a four-limb problem, and it's been done at least once. Grabbing the rungs with all limbs will make it easier.

Replacing a pump is an interesting problem. I don't think anyone has yet demonstrated complex part replacement in an unstructured environment. That's probably the least well developed part of the problem.

Building a suitable machine will be expensive. Each Boston Dynamics robot has cost upwards of $20 million. Even the Willow Robotics machine is over $100K per unit, and it's just two arms on a wheeled base. This thing has to have roughly human dimensions and be self contained. Petman isn't self-contained; it needs external power.

Yet, looking at the problem, I can see how to approach it. Modern control theory is good enough. Machine learning and vision processing are good enough. Simulators are good enough to allow debugging in simulation. Enough people know this stuff that the job can be staffed.

Re:Why little to no teleoperation?

[Mr. Freeman]

You'd think so, yes. Of course, DARPA isn't so much about making useful things as it is about spurring research that can be used to later create useful things. I suspect that's why they are going with a competition that no one will be able to complete in the next 3 years, and certainly not in a reasonable time frame.

This has happened before. It took two years before any competitor was able to even finish the off-road DARPA grand challenge, and that was just driving. A bi-pedal robot is orders of magnitude more complicated.

I predict that they're going to have to start each section of the race separately. They'll put the bot in the tractor, start the race. Then they'll move the bot to the door, start it again, etc. No way in hell anyone is going to be able to do these tasks in sequence. Some they probably won't be able to do at all.

Re:That's ambitious

[Missing.Matter]

Machine learning and vision processing are good enough.

I disagree on both counts. Machine learning is very good at specific problems. Current best techniques to train classifiers need millions of samples of ground truth training data, which is incredibly expensive to produce. Most classifiers like this make several assumptions about the training data, namely that it is representative of a stable underlying unknown distribution, which of course is hardly true. The best uses for machine learning techniques are for very specific applications where this assumption can be enforced, like optical character recognition. Machine learning has amazing applications, but it's hardly a field where I would say the results are "good enough," especially for most outdoor mobile robotics applications, like this DARPA project.

Same goes for vision processing. The reason most robots today use LIDARs as their primary "vision" sensor is that image processing is very slow and not very good. With a 3d LIDAR you can do mapping, object detection and recognition, people detection, grasp planning, obstacle avoidance, etc. all very easily in polynomial time. Most of these tasks are computationally intractable using images, if they're possible at all (for example, night time operations).