July 20, 2010      

Early this year, the Defense Advanced Research Projects Agency (DARPA) released a Broad Agency Announcement, or BAA, calling for proposals for a three-phase research project to solve hardware and software problems related to robotic manipulation. The result was the Autonomous Robotic Manipulation (ARM) program, led by DARPA program manager Dr. Robert Mandelbaum. Mandelbaum is also the manager for other well-known DARPA robotics programs, such as the BigDog quadruped and the LANdroid ground robot.

Manipulation is a major challenge in robotics, and this is not the first robotic arm program that DARPA has funded. The “Luke arm,” for example, is a direct neural-driven prosthetic arm developed by DEKA Research and Development under DARPA funding. But many manipulation efforts focus on direct-driven or teleoperated systems, not autonomous manipulation. The ARM program is designed to address this fact.

The ARM program consists of both a hardware and software track. For the software track, teams are provided with a standardized hardware platform. DARPA selected Barrett Technology Inc. and re2 Inc. to develop the platform to be used for the software development track. Barrett is providing its seven-degree-of-freedom (DOF) “WAM” arms and three-fingered tactile sensing grippers-a popular platform for academic and research development-and re2 is acting as a system integrator, building a unified platform that includes a “head” containing infrared and video cameras, a stereo microphone, and a pan-tilt unit. Using this standardized hardware, developers will write and test autonomous manipulation software to carry out three separate tasks.

The hardware teams must complete the same tasks, competing against the software teams working with the provided platform, but must build their own multi-DOF manipulators to integrate with the WAM arm. Code from the software track will be loaded on the hardware track platforms to test cross-compatibility.

The Trial Scenarios
The tasks test manipulation capabilities, which historically have been the most difficult for robotic systems to perform. This includes grasping and manipulating nonrigid and large irregular objects, as well as manipulating two objects with respect to one another. To test these, DARPA has set up trial scenarios that mimic many of the typical applications for bomb disposal or search-and-rescue robots.

For the first task, the robot must open a gym bag, search through its contents, and identify and remove a revolver. In the second, to simulate a rescue effort from rubble, the robot must pick up and move a large rock. In the final task, the robot must load a shell into a mortar. Each phase will require teams to perform increasingly more tasks. DARPA will downselect to the most successful teams at the end of each phase.

Though DARPA is providing developers with powerful computer-vision hardware, vision and learning are only a tangential focus of this program. For tasks such as the revolver retrieval, teams are provided with 3D CAD models of the object they are trying to locate and remove. To some extent, robotic “common sense” is required; for example, one task may involve manipulating a screwdriver that is longer or wider than the model provided to the team. However, autonomously “learning” and reliably identifying an object is not a major objective-for now.

The hardware teams, which will be developing new manipulators to compete against the Barrett hand on the standard platform, are part of a compelling effort to drive low-cost robot design. Advanced manipulators are often expensive, and less expensive designs must rely on single-DOF grippers, not “hands” with discretely operated fingers. Though the software track may generate many interesting new capabilities for robotic manipulation, the hardware track, if successful, could result in immediate adoption in a number of commercial and industrial applications in a space that urgently needs cheap, capable solutions. It is also an interesting move for a government agency to be funding a low-cost design effort, not just pushing new technical capability.

Companies in the Running
Attendees at the DARPA program industry day-a day-long introduction to the program and an opportunity to ask questions before doing a proposal-give some indication as to who may be in the running. Companies such as iRobot Corp., QinetiQ North America, Meka Robotics, and Boston Dynamics will be competing with larger, primary contractors like Boeing, Lockheed Martin, Raytheon Sarcos, and SAIC. Academic labs such as the Applied Research Laboratory at Penn State, Carnegie Mellon, and MIT may also be submitting proposals.

Prospective participants submitted their proposals in late March for an initial round, though proposals may be submitted (with reduced likelihood of funding) through August 2010. The contract award date is not currently known. The first 15-month-long phase will allow six teams to compete; the third and final phase will consist of only two teams.