Section: New Results

Multi-robot and Crowd Motion Control

Advanced multi-robot control and estimation

Participant : Paolo Robuffo Giordano.

The challenge of coordinating the actions of multiple robots is inspired by the idea that proper coordination of many simple robots can lead to the fulfillment of arbitrarily complex tasks in a robust (to single robot failures) and highly flexible way. Teams of multi-robots can take advantage of their number to perform, for example, complex manipulation and assembly tasks, or to obtain rich spatial awareness by suitably distributing themselves in the environment. Within the scope of robotics, autonomous search and rescue, firefighting, exploration and intervention in dangerous or inaccessible areas are the most promising applications.

In the context of multi-robot (and multi-UAV) coordinated control, connectivity of the underlying graph is perhaps the most fundamental requirement in order to allow a group of robots accomplishing common goals by means of decentralized solutions. In fact, graph connectivity ensures the needed continuity in the data flow among all the robots in the group which, over time, makes it possible to share and distribute the needed information. We gave two contributions in this field: in the first one [35], we proposed a decentralized exploration strategy for a team of 3D agents able to guarantee exploration of a finite space in a finite amount of time while coping with the constraints of a connected sensing/communication graph for the robot group against sensing/communication constraints (limited range, occluded line-of-sight), and of obstacle and inter-robot collision avoidance. The strategy exploits a suitable state machine for assigning dynamic roles to the agents in the group for allowing completion of the exploration in finite time. Second, in [28] we studied how the choice of a leader agent in a leader-follower scenario could affect the performance of the group when tracking a desired formation (shape and gross motion). The proposed strategy allows selecting the “best leader” online as a function of the current group state (relative positions and velocities) and of the group topology (assumed connected). By cycling among several connected topologies during motion, we could show that our proposed leader selection algorithm provides the best performance among other possible choices (including the random one) while coping with the constraint of a connected (but possibly time-varying) topology.

These works were realized in collaboration with the robotics group at the Max Planck Institute for Biological Cybernetics, Tübingen, Germany, and the RIS group at Laas in Toulouse.

Rigidity-based methods for formation control

Participants : Fabrizio Schiano, Riccardo Spica, Andrea Peruffo, Paolo Robuffo Giordano.

Most multi-robot applications must rely on relative sensing among the robot pairs (rather than absolute/external sensing such as, e.g., GPS). For these systems, the concept of rigidity provides the correct framework for defining an appropriate sensing and communication topology architecture. Rigidity is a combinatorial theory for characterizing the “stiffness” or “flexibility” of structures formed by rigid bodies connected by flexible linkages or hinges. In a broader context, rigidity turns out to be an important architectural property of many multi-agent systems when a common inertial reference frame is unavailable. Applications that rely on sensor fusion for localization, exploration, mapping and cooperative tracking of a target, all can benefit from notions in rigidity theory. The concept of rigidity, therefore, provides the theoretical foundation for approaching decentralized solutions to the aforementioned problems using distance measurement sensors, and thus establishing an appropriate framework for relating system level architectural requirements to the sensing and communication capabilities of the system.

In the recent past, we have proposed a decentralized gradient-based rigidity maintenance action for a group of quadrotor UAVs [10]. By starting in a rigid configuration, the group of UAVs was able to estimate their relative position from sole relative distance measurements, and then use these estimated relative positions in a control action able to preserve rigidity of the whole formation despite presence of sensor limitations (maximum range and line-of-sight occlusions), possible collisions with obstacles and inter-robot collisions. This (rigidity-based) control/estimation framework has now been extended to the case of bearing rigidity for directed graphs: here, rather than distances the measurements are the 3D bearing vectors expressed in the local body-frame of each agent. The theory has been extended to the case of 3D agents evolving in ${ℝ}^3\times {𝒮}^1$ by proposing a decentralized bearing controller/localization algorithm that only requires one single distance measurement (among an arbitrary pair of agents) for a correct convergence [72]. The proposed algorithm ensures stabilization towards a desired bearing formation, and allows for the possibility of actuating the motion directions in the null-space of the bearing constraints (that is, collective translations in 3D, expansion/retraction, and coordinated rotation about a vertical axis).

The need of a single distance measurement (for fixing the formation scale) has also been relaxed in [73] where an active scale estimation scheme has been proposed for allowing the (distributed) estimation of the various inter-agent distances online by processing the measured bearings and the known agent ego-motion (body-frame linear and angular velocities). Finally, we have also proposed an extension of the “distance” rigidity maintenance controller proposed in [10] to the case of bearing measurements (and bearing rigidity), by considering the typical sensing constraints of onboard cameras, that is, limited range, limited field of view, of possible mutual occlusions when two or more agents lie on the same line-of-sight. This work has been experimentally validated with 5 quadrotor UAVs, and has been submitted to ICRA'2017.

These works were realized in collaboration with the RIS group at Laas, Toulouse, and with Technion, Israel.

Cooperative localization using interval analysis

Participants : Ide Flore Kenmogne Fokam, Vincent Drevelle.

In the context of multi-robot fleets, cooperative localization consists in gaining better position estimate through measurements and data exchange with neighboring robots. Positioning integrity (i.e., providing reliable position uncertainty information) is also a key point for mission-critical tasks, like collision avoidance. The goal of this work is to compute position uncertainty volumes for each robot of the fleet, using a decentralized method (i.e., using only local communication with the neighbors). The problem is addressed in a bounded-error framework, with interval analysis and constraint propagation methods. These methods enable to provide guaranteed position error bounds, assuming bounded-error measurements. They are not affected by over-convergence due to data incest, which makes them a well sound framework for decentralized estimation. Ongoing work focuses on position uncertainty domain computation in image-based UAV localization [63], and its extension to cooperative localization in a multi-UAV fleet.

Numerical Models of Local Interactions during Locomotion

Participants : Julien Bruneau, Panayiotis Charalambous, David Wolinski, Julien Pettré.

The numerical models of local interactions are core components of reactive navigation techniques (which allows a robot to avoid dynamic obstacles) and of microscopic crowd simulation algorithms (which allows to simulate a crowd motion as a collection of agent trajectories). We have pursued our efforts to design local models of interactions which capture humans pedestrian behavior, to simulate how they adapt their trajectory so as to perform interactions with their neighbors [12]. This year, our efforts were focused on the simulation of grouping behaviors [39], and mid-term strategies human set to perform energy-efficient sequences of successive avoidance adaptations [24]. These two situations deal with complex situations of interactions, where several interactions of different kinds need to be combined to compute agents trajectories. For example, when moving in groups, agents have to keep close to the other members of their group while they should not collide with them, as well as they should avoid collision with any other agent or obstacle out of this group.

We also revisited the foundation of velocity-based models of local interaction for collision avoidance. Using a velocity-based model, a collision-free motion is computed for one agent by extrapolating the future motion of neighbor agents with respect to their current position and velocity. From this information, each agent can deduce the set of velocities (called admissible velocities) that lead to a collision-free motion in the near future. The extrapolation is generally simply based on a linear extrapolation of the future position along the current velocity vector. This is simplistic as it assumes that the current velocity vector is representative of the future motion, while it is often false when, for instance, the agent is currently performing adaptations due to ongoing collision avoidance, or when the agent is following a curvy path. To improve the accuracy of motion prediction and the resulting simulation, we have introduced a probabilistic representation of future position, that can be computed from a set of context elements such as the layout of the environment or the agents past motion [42]. We demonstrate in this work the high impact on the level of realism of resulting simulations. This work is implemented in the WarpDriver software (see Section 6.7).

Finally, we address applications of our simulators to the Computer Animation. Crowd simulation agents generally have a simplistic geometrical and kinematics models, typically, an oriented 2D circle moving on a flat surface. In Computer Animation, an animation of a crowd of 3D realistic characters can be computed on top of the agents simulation by computing their internal joints trajectories so as to perform walking motion along computed agents trajectories. However, the discrepancies between the 2D model of agents and 3D full body characters may result into residual collisions between character shapes. In this collaboration with the Mimetic team, we demonstrate that simple secondary animations for characters, such as local shoulder motions, can be efficiently triggered to camouflage those artefacts, with a very low computational overhead [29].

Motion Planning for Digital Characters

Participant : Julien Pettré.

Motion planning is an important component for agents and robot navigation and control, providing them the ability to perform geometrical reasoning over their environment to transform a high-level distant goal in their environment into a sequence of local motions and sub-goals to reach. This year, we have been involved into two collaborations dealing with motion planning. First collaboration was with the University of Utrecht in the Netherlands. We have proposed a method to evaluate and compare various environment decomposition techniques [74]. Environment decomposition is an important step to perform navigation planning in large static environments. Second collaboration was with the University of North Carolina in Chapel Hill (see Section 9.4.1.1). We have coupled a contact planner for virtual characters with ITOMP, a motion optimization technique to achieve complex motion in cluttered environment [69].