Notes:
- Everybody using the Atlas robot had stability problems. Nobody seemed to be outside of static stability (base of support under CG) let alone outside of ZMP stability (base of support under CG modified to account for momentum.) Any loss of balance meant a fall. They're probably using the basic DLL that Boston Dynamics provided to get teams going. (One team admitted this.) That has functions "walk slowly", and "stand maintaining balance using only joints from the hip and below", plus direct control of joints. Boston Dynamics has demonstrated much better locomotion control than that, but the teams didn't have access to it. (Boston Dynamics supplied the Atlas robots and thus could not compete.) IHMC was working on balance recovery [4] but didn't have it working in time.
- This is the new Atlas robot, unveiled in January 2015. It's still mostly hydraulic, but it's now self-powered, with a big battery backpack. I think there's also a hydraulic accumulator in there, so they can recover some energy when running, but nobody at the event went fast enough to gain from that. Mechanically, that machine is far more capable than it appeared to be in the competition. It's quite capable of picking up objects from the ground, and getting up if it falls. Nobody did either in the competition. The teams only got these new robots around February, so they really haven't had the time to develop software to use them effectively. Right now, the hardware is much better than the software.
- Everybody was doing their manipulation by almost pure teleoperation. As soon as DARPA introduced delays into the data links (which they do to penalize teleoperation), the robots spent most of their time stopped. The operators were stuck waiting for enough data to get through to update their screens. Not much autonomous operation here. Here's the winning team video, sped up about 30x.[2] Note that although it's capable of walking, it was run in kneeling wheel mode almost all the time. Only for the stair climbing did it walk.
- Nobody had combined manipulation and locomotion working at all. That's why getting out of the vehicle was so hard - it takes coordinated leg and arm movements.
- The CMU team had a partially tracked base, and the NASA/JPL team had a "monkey" robot with rollers at knees and elbows. They both did reasonably well. If you're going to teleoperate, a dumber, statically stable robot is easier to deal with. Of the top 3 finishers, only one used an non-wheeled robot, an Atlas. Here's their first run, with two falls.[3] Again, this is a 30x speedup. The penalty for a fall or a need for repair was only a 10 minute penalty, not disqualification.
- The tasks were dumbed down twice during the project. The original specs called for the robots to have to replace a pump. The final competition only required turning a valve and cutting a hole in very soft material with a power drill. Basically, the tasks were reduced to the point that they were hard, but not so hard as to lead to total fail.
- Schaft did better in the first round back in 2013, but after Google bought them, they were pulled from the competition. Google, remember, now owns Schaft, Boston Dynamics, and several other robotics companies.
I look forward to the next competition. If work continues, in a year or two the results should be far better.
[1] https://www.youtube.com/watch?v=27HkxMo6qK0 [2] https://www.youtube.com/watch?v=q_yfvlplJe0 [3] https://www.youtube.com/watch?v=vDXnlgNkRl4 [4] https://www.youtube.com/watch?v=soPM7LddLY8