matteosodano
commented
4 years ago Multi-contact heavy object pushing with a centaur-type humanoid robot: planning and control for a real demonstrator - Polverini, Laurenzi, Hoffman, Ruscelli, Tsagarakis A NLP multi-contact planner together with a multi-contact controller is proposed to define contact points and joint trajectories. The former plans the CoM positions, the contact positions and the contact forces, solving a NLP problem using the robot centroidal dynamics under quasi-static assumptions. The latter tracks at best the output of the multi-contact planning layer, producing reference joint positions and feedforward torque solving a hierarchical inverse kinematics problem. To allow contacts with walls and ground, superquadric functions are used to model the environment.
Gait and trajectory optimization for legged systems through phase-based end-effector parametrization - Winkler, Bellicoso, Hutter, Buchli A trajectory optimization problem for legged locomotion that automatically determines gait sequence, step timings, footholds, swing leg locomotions, trunk pose and contact forces over non-flat terrain is formulated. This allows to efficiently generate complex highly-dynamic motions. Initial and final states, total duration of the trajectories and the amount of steps per foot are provided. The foot motions and the ground reaction force in each phase (swing or contact) is represented by either a constant value or a sequence of cubic polynomials. The robot is modeled by means of the CD. The kinematic model is conservatively approximated by keeping each foot inside its range of motion, i.e. it is assumed that joint limits are not violated if every foot lies inside a appropriately defined cube (dependent on the current CoM position).
Dynamic locomotion through online nonlinear motion optimization for quadrupedal robots - Bellicoso, Jenelten, Gehring, Hutter The motion planning framework consists of two main modules, namely a CoM motion planner and a foothold generator, both of which are separately solving an optimization problem. The plan is computed online and executing them in a MPC-like fashion, i.e. continuously run the motion optimizer initialized with the most recent state of the robot to update the motion plan that is tracked by the whole-body motion controller. The motion plan that they are optimizing is described over a sequence of support polygons which depends on the type of gait that is being executed. Each coordinate of the CoM motion plan is seen as a sequence of quantic splines: for each component of the CoM, there exists at least one spline for each support polygon; two splines are used for full-flight phase. The motion optimization relies on a SQP framework to solve the nonlinear ZMP constraints. They exploit the complete dynamical model of the robot, and run the algorithm online. The CoM height is variable.
Bellicoso, Jenelten, Gehring, Hutter - Dynamic Locomotion through Online Nonlinear Motion.pdf
Online walking motion generation with automatic foot step placement - Herdt, Diedam, Wieber, Dimitrov, Mombau, Diehl Obtain fully automatic placement of the foot steps: a reference speed is given to the robot, which can be modified any time, and according to this reference speed and the current state of the robot a safe foot step placement is decided. Even in case of external perturbation (or if the reference is not realizable) the generated walking motion is always entirely safe and stable: the reference speed is tracked only as much as possible within the limits of stability. The CoM is constrained to move on a horizontal plane at height h from the ground. Therefore, a linear MPC control scheme is used: having short computational times, online use is allowed. They want to minimize the jerk input, the tracking error from the reference speed and the tracking reference error of the ZMP, satisfying contraints on the CoP and on the foot placement.
Trajectory generation for multi-contact momentum-control - Herzog, Rotella, Schaal, Righetti Using the centroidal dynamics, the authors presented an approach to control contact forces and momentum for legged robots. CoM positions, contact forces and momenta are obtained solving a quadratic optimization problem imposing some linear constraint. Feedback gains are generated from a LQR design, linearizing the nonlinear momentum dynamics around the optimized trajectory. A whole-body controller computes the necessary (and feasible) joint torques and leg motions. The gait is imposed before-hand, for example by an higher-level planner.
Herzog, Rotella, Schaal, Righetti - Trajectory Generation for Multi-Contact Momentum-Control.pdf
Real-Time motion planning of legged robots: a model predictive control approach - Farshidian, Jelavic, Satapathy, Giftthaler, Buchli Real-time, constrained, nonlinear MPC for the motion planning of legged robots. The algorithm continuously re-optimize state and control input trajectories based on dynamic programming (in particular, an SLQ-like algorithm) using a multi-processing scheme for estimating value-function in its backward pass. Performance are demonstrated for trotting.
Simultaneous contact, gait and motion planning for robust multi-legged locomotion via mixed-integer convex optimization - Aceituno-Cabezas, Mastalli, Dai, Focchi, Radulescu, Caldwell, Cappelletto, Grieco, Fernandez-Lopez, Semini The authors present an approach for simultaneously planning contacts and motions for multi-legged robots on complex terrains based on mixed-integer convex programming. Modeling the robot using CD, automatic gaits are computed incorporating the gait sequences as decision variables. Convex representation of friction cones, torque limits and gait planning constraints are incorporated. The global optimum is found because of the convexity of the problem.
Hybrid direct collocation and control in the constraint-consistent subspace for dynamic legged robot locomotion - Pardo, Neunert, Winkler, Grandia, Buchli Instead of considering the full robot dynamics and constraining the motion, in \cite{pardo2017hybrid} the dynamical model is projected into the null space of the constraints. The search of feasible and stable trajectories for states and controls is restricted to a constraint--consistent subspace, eliminating contact forces from the set of decision variables. This method requires to specify the contact sequence in advance, nevertheless it is possible to achieve complex dynamical motions, such as walking, trotting and jumping.
Let's find some nice papers on OC used to plan trajectories for biped and quadruped robots.