ADAMANT: Adaptive Manipulation for Tasks

Robots will play an important role in NASA's upcoming missions to the Moon and beyond. More than just remote sensors, they will be expected to manipulate their environment in a complex and useful way - carrying objects, using tools, and assisting the crew with various physical activities. NASA has been developing world-class dexterous end effectors for years. Unfortunately, developing software to fully utilize such hands is very challenging. Grasping strategies tend to be highly dependent on object models and localization, or reliant on a good connection to an operator. As any of these deteriorate, even simple grasping of known objects becomes unreliable. The environment or the object's intended use can influence how to grasp it. The best way to pick up a tool will depend on whether it is to be transported to another location, handed to a crew member, or used as a tool.Previously with NASA, TRACLabs developed robot control software called CRAFTSMAN that includes trajectory generation, simple action-sequencing capabilities, and a method for parameterizing, encoding, and visualizing task descriptions. CRAFTSMAN supports robot-independent task descriptions, but grasp strategies are still robot-specific open-loop waypoint sequences, subject to the problems listed above. In this work, we propose to extend CRAFTSMAN to handle grasping as a task-informed behavior, using sensor data and object models when possible to identify grasp sites. This new system, called ADAMANT (ADAptive MANipulation for Tasks), will help an operator to determine the best option for acquiring an object. The result will be a robot grasping interface that is more intuitive to use than current technology and will produce more robust robot behavior. This will reduce the cognitive load on remote robot operators by eliminating the need for run-time manual adjustments. By removing the details of grasp strategy from high-level planning, the design of action sequences will also become easier.