What is the DARPA Robotics Challenge?

 Disaster response situations like nuclear reactor leaks, mine collapses, and deepwater oil spills are extremely dangerous for humans, yet human experts are typically the only ones capable of mitigating the hazard.  The DARPA Robotics Challenge (DRC) is designed with the aim to develop hardware and software which would make putting a robot in a disaster situation instead of a human, not only viable, but the most obvious solution.

However, there are many difficulties which need be addressed before such a solution is a reality.  While robotic capabilities are rapidly advancing, hardware and software to perform such complex tasks as walking, driving, and manipulating objects in complicated environments with limited connection between robot and user is not yet available. The DRC addresses the problem from multiple angles.  Track A teams are working on designing robots which are able to do all the difficult physical tasks.  The University of Washington team is competing as a track B team in the virtual robotics competition to design the software which will enable the DRC robot (see image below) to carry out the requested tasks.

The UW team is lead by Professor Emanuel Todorov.  He is joined by many other experts and students working on the varied aspects of the project (controller design, computer vision, perception, human-computer interaction,....the list goes on).  Professor Dieter Fox, co-PI, is leading efforts in the computer vision and perception areas. If you are interested in working with us, we'd love to hear from you!

DRC Robot from Boston Dynamics

Tasks of the competition 

The competition involves four tasks to be performed outside in daylight.

  1. Robot gets into, starts, stops, and gets out of a vehicle. 
  2. Robot drives a vehicle on a road with "sparse large obstacles".
  3. Robot walks on a 100m obstacle course with obstacles and varied terrain.
  4. Robot connects a hose or cable to a wall socket.

The competition taking place June 10-24, 2013 will involve 5 runs on each task. The initial configuration, communication bandwidth and latency, and contact friction (wet/dry) will be varied. It is thought that the tasks will take about 10 minutes with a 30 minute time limit. 

Performance will be scored according to the following critereon:

  • Task completion: did you succeed in doing task?
  • Time: how long did it take?
  • Data communications volume in both directions between robot controller and operator station.
  • Maximum absolute acceleration of center of mass: an attempt to score whether you fell down, and how hard you hit the ground.
Prior to the full competition, there will be a qualifying round May 1-15.  There will be no constraint on time, nor will there be any penalization for data volumes.  The teams must demonstrate successful completion of walking with stairs and other obstacles and also a manipulation task to pick up and move a small object.

Our approach: end-to-end optimization

We plan to use a unified approach to automatic control, human supervision, and machine perception based on numerical optimization. At the core of our approach is optimal control – in particular trajectory optimization with respect to a detailed physics model which runs fast enough to enable model-predictive control (MPC). The application of MPC to a full humanoid, without any model reduction, is unprecedented. Our recent work has made this possible through several coordinated breakthroughs: a new full-featured physics engine which we have designed specifically for control purposes and which runs orders-of-magnitude faster than real-time taking advantage of multi-core processors [1]; new mathematical formulations of the physics of contact which yield efficient algorithms whose output is differentiable and yet corresponds to hard contacts [2, 3]; improvements to our trajectory optimization methods which were already state-of-the-art [4]; and novel formulations of the optimal control problem that facilitate synthesis of complex movements [5, 6]. The ability to apply MPC to a full humanoid will be a significant advantage in the present context, because it ”invents” the movements on the fly and thereby can recover from pretty much arbitrary disturbances as well as changes in the environment or task objectives. We will also take advantage of more traditional offline controller designs, which will be automatically combined and exploited by the MPC machinery.

DRC Robot model