CAMP

An effective way of dealing with the complexity of robots operating in real (uncertain) environments is the paradigm of feedforward/feedback or planning/control: in a first step a suitable nominal trajectory (feedforward) for the robot states/controls is planned exploiting the available information (e.g., a model of the robot and of the environment). This step is usually executed offline and can take into account presence of constraints (e.g., collision avoidance, limited actuation) and optimality w.r.t. metrics of interest (e.g., time, energy). An open-loop execution of this planned trajectory would, however, fail in most practical cases because of the unavoidable approximations and uncertainties affecting the model used for planning. Therefore, the planned trajectory is in practice robustly executed by making use of a motion controller that closes the loop between planned and actual motion, providing a robustness layer against all the unmodeled effects that could not be considered at the planning stage. This sequential composition, treating planning and control as separated components, has however several shortcomings: on the planning side, modern planners can generate feasible and globally optimal paths for high-dimensional systems and complex constraints/environments. However, they do not typically consider (and, even less, exploit) the unavoidable presence of a runtime feedback controller that will actually execute the plan. On the control side, a large number of adaptive or robust (e.g., H-infinity or passivity-based) control schemes have been proposed over the decades for delivering a good level of robustness against uncertainties/disturbances. However, these approaches are mostly local and, besides robustness, can hardly tackle other constraints/requirements such as feasibility (e.g., limited actuation, obstacle avoidance), performance and global optimality, which can otherwise be effectively handled by traditional planners. While there has been an effort in proposing robust planners or more global controllers (e.g., Model Predictive Control), a truly unified approach that fully exploits the techniques of the motion planning and robust control communities is still missing and the existing state-of-the-art has several important limitations, namely (1) lack of generality, (2) lack of computational efficiency, and (3) poor robustness. All these shortcomings are a major limiting factor in the autonomy and decision-making capabilities of robots operating in all those complex scenarios (real-world conditions, non-negligible effects of the uncertainties, fast dynamics) which are instead the typical conditions in which future robots are expected to operate. Therefore, the ambition and scientific goal of CAMP is to (1) develop a general and unified intrinsically-robust and control-aware motion planning framework able to address all the above-mentioned issues, and to (2) demonstrate the applicability of this new framework in a real-world challenging scenario involving a pick-and-place task for a 7-dof robot arm, and a cooperative manipulation task for a mobile robot and aerial manipulator. These tasks are purposely chosen for showcasing at best the benefits of the proposed planning framework since accuracy (implying low execution uncertainty) and execution speed (resulting in a significant impact of unmodeled effects and disturbances) are crucial. The consortium of CAMP regroups three teams with a consolidated (and complementary) competence in planning, estimation and control. If successful, the project will have concrete benefits for any robotic application in which uncertainties play a significant role in all areas of robotics.