Thursday Practice: First Impressions
It all started with practice sessions on Thursday - which turned out to be a snoozer. It was clear that the teams were still getting used to the course and had very little practice. As we learned later on, there were also network issues that had not been fixed yet - this was a big issue for the teams since they were all using tele-operation to operate the robots. DARPA has specified that shaping the available bandwidth would be a part of the trials, presumably to encourage teams to build as much autonomy into their robots as possible. With the limited lead time for the contest though, all the teams were still operating using teleop and had to account for the changes in the network - the network switched between 100 Kb/s and 1 Mb/s every minute. A 1-second latency was also added in.
I also had my first look at ATLAS: the Boston Dynamics built robot that was provided to several teams for the contest. Clearly, the robot is a beast - its heavy, tall and noisy (it carries its own hydraulic equipment on its back). It has to be tethered for cooling water to flow into the robot. Boston Dynamics provides a black-box walking library with the robot which most teams were still using (except apparently a couple of notable exceptions: IHMC and WRECS). The robot also drips and loses hydraulic oil - one reason that the Miami speedway was a good choice of venue. Its arms are only 6-DOF with a limited workspace creating a lot of problems for teams when they need to do tasks like opening valves or climbing the ladder.
The robot is equipped with a couple of options for hands but, surprisingly, several teams opted to put hooks on the robot - these were especially useful for tasks like opening doors and climbing a ladder. The rules of the competition required that the teams had to carry any tools and alternate hands or grippers they would use throughout the competition in all the rounds - e.g. if they wanted to use a hook for one task and then a gripper for another task, they would have to carry both on the robot at all times. A popular gripper on the robots seemed to be the Robotiq gripper - a heavy-duty versatile gripper developed for research and industrial uses.
The day clearly belonged to Team SCHAFT though. A Japanese company originally out of the JSK lab in the University of Tokyo (the same lab that was also a part of the PR2 Beta Program) build the robot based on the design of the HRP-2 robot. SCHAFT is a beautiful robot, clearly the result of the years of experience that lab has had in building humanoid robots. I saw it walking over the rough terrain, planning and taking two steps at a time. I have seen several robots walking but clearly this robot was in a league of its own - it was the smoothest walking robot I had ever seen. The arms are also mounted in what, at first sight, seems like a very weird configuration. However, as I learned, the design of the kinematics for the arms was refined using a workspace analysis. This meant that the robot is able to reach and operate with its arms within a very big workspace in front of the body. Clearly, Team SCHAFT has given much more thought to this design issue than Boston Dynamics.
There were a variety of teams, a variety of robots and a variety of tasks for the robots to solve. There was one thing in common with a lot of the teams though - most of them were using ROS. A look at the introductory videos for most teams had screenshots from Rviz (the ROS visualizer) - making it a proud moment indeed for Dave Hershberger (ex-Willow and now at SRI with me) who wrote the latest version of Rviz.
Friday Competition: Off to the Races!!
Thursday may have been a snoozer but come Friday, the robots were off to a strong start. It was clear that the network issues that may have existed the day before were now gone. Teams were opening doors, walking over terrain and working at all the tasks much more confidently than the day before. Team SCHAFT was clearly the pick of the lot though, bursting out of the gates and staying ahead throughout. However, the going overall was slow. It took close to half an hour for each team to get through individual tasks. My lasting impression of this day was of watching one robot take one step, walking over to another task to see a different robot move a little and then coming back to see the first robot take another step.
The teams had opted for a conservative approach to the contest. The robot operators were taking their time examining the terrains - mainly in 3D using scanning laser sensors. The cognitive decisions were all being made by the humans - they decided where the robot would put its feet and where they would move their arms. For me, this was slightly disappointing - having worked with the PR2 on autonomy for so long, I was hoping to see more autonomy in the task. Given what was at stake though, (only the top 8 teams would qualify for continued funding), it was understandable.
As the competition progressed, it was also clear that DARPA had made some of the tasks just that much easier. The wood in the debris competition seemed to always be arranged a certain way. The robots were no longer expected to get in and out of the car. The valves were apparently easier to turn. Some tasks got harder though - the wind at the venue kept closing the doors on the robots.
Saturday: We have a winner!!
Saturday saw some of the best walking on display as team SCHAFT tackled the terrain and the ladder. It was clear that Team SCHAFT was way ahead of the other teams. They had the railings removed from the ladder and then proceeded to climb to the top without needing a handhold. The ATLAS robots did not fare as well on the ladder although Team WRECS managed to climb up a long way, using the hands to hold on to the steps as they walked up.
SCHAFT ran away with the competition, scoring 27 out of a total of 32 points. The next best team, IHMC, got to 20 points total. Three teams placed in the top eight using non-ATLAS robots: SCHAFT, Tartan Rescue and RoboSimian. Tartan Rescue had an interesting robot - using a combination of tracks and limbs to create a hybrid robot that aimed to combine the best of both. RoboSimian also used a combination of wheels and limbs and a statically stable walking gait. The ATLAS robots placed 5 teams in the top eight: IHMC, MIT, TracLabs, WRECS, Trooper.
Where do we go from here?
The trials exceeded expectations. 4 teams scored more than half the total number of points possible. The teams clearly did well at the tasks in the trials themselves, establishing a benchmark for the state of the art. The improvements in walking were incredible to see. The new hardware solutions coming out from the trials are revolutionary, ranging from the large powerful robots like SCHAFT and ATLAS to the smaller systems like THOR. The software and systems on display were also very impressive.
It is less clear how the skills on display would translate to an actual search and rescue scenario. The tasks were very structured and some of the teams clearly programmed to the tasks. More randomization would help in making the tasks more pertinent. Its also clear that more autonomy will become essential when the robots are required to venture into more unstructured scenarios, situations where software frameworks like MoveIt! will become more useful.
Walking robots may not be the most practical solution in the future- most current rescue robots are tracked. Search and rescue also often involves moving into cramped spaces - something that the giant robots like ATLAS and SCHAFT will not be capable of doing. The robots will have to carry their own power source into such scenarios as well - it is rumored that DARPA may require ATLAS to carry its own cooling. I am hoping to see more of the electric robots coming to the fore as the contest progresses - although the hydraulically powered ATLAS is very capable, a giant leap on the electric side would help in driving robots in other domains. The best looking robot in the trials - NASA Johnson's Valkyrie - did not do very well since the team was so focused on building hardware that they did not get the time to test out their robot. It would be great to see this robot do well in the next year. THOR just missed out on the top eight - their robot is smaller and leaner than the others and I hope they come back with more practice the next year.
The venue for next year's competition has not been finalized. The Miami Speedway seemed like a good venue although the heat and humidity may have affected several robots - especially since a lot of the robots are black in color. The Speedway does have garages (where the teams were staged) and is used to having a lot of dripping oil around. The constant thunder from fighter jets flying overhead only added to the overall atmosphere of the trials. Last week's event was a first step in the road to robots helping in search and rescue. It will be a very long road - but its definitely one worth pursuing to the end.