Short State-of-the-Art (SoA) on Control-based sample methods

[EnricoMingo]

As title, lets prepare a short soa on control based methods applied to problems similar to the gap_crossing task. Put here references, for each reference you also write a short summary on the method.

[matteosodano]

We have found two articles related to the use of Motion Primitives in humanoids locomotion. We think that the first one is very promising for our work, while in the second one there are some interesting ideas that can be exploited.

Whole-Body Motion Planning for Humanoids based on CoM Movement Primitives - Marco Cognetti, Pouya Mohammadi, Giuseppe Oriolo It introduces a randomized planner that builds a tree in configuration-time space by concatenating feasible, collision-free, whole-body motions realizing a succession of CoM movement primitives. Each primitive has a given duration and may specify trajectories for other points of the robot in addition to the CoM (e.g. swinging foot). A precomputed catalogue of CoM movement primitives is considered (they use static steps, dynamic steps and non-stepping motion, but other primitives can be considered). A hybrid (algebraic + ODE) model for the robot is considered. A planner selects the CoM primitive and computes the relative joint velocities using the Inverse Differential Kinematics. It also checks that collisions do not occur and position and velocity limits are not violated. Cognetti, Mohammadi, Oriolo - Whole Body Motion Planning for Humanoids based on CoM Movement Primitives.pdf

Related works:

• Whole-Body Planning for Humanoid along deformable tasks – Cognetti, Fioretti, Oriolo
• Humanoid Whole-Body Planning for Loco-Manipulation Tasks – Ferrari, Cognetti, Oriolo
• Task-Oriented Whole-Body Planning for Humanoids based on hybrid motion generation – Cognetti, Mohammadi, Oriolo, Vendittelli

Using Motion Primitives in Probabilistic Sample-Based Planning for Humanoid Robots – Hauser, Brelt, Harada, Latombe A feasible space F is defined according to some constraints, such as equilibrium and collision avoidance (except at contacts with the ground). They assume that the robot moves only by taking steps. Primitives are defined as “high quality” steps according to some criteria (e.g., minimize path length, torque, energy, etc.). In the first approach, they assume to know the robot initial configuration and its stance as well as its final configuration with stance. Then, some intermediate configurations between q_init and q_final are chosen; it is checked that they belong to F and, if they do, they are added to a tree (otherwise they will be discarded, and another sample will be generated). A list [q_init q_1 … q_final] is obtained. For each pair (q_i, q_i+1), it is checked if there exist a primitive that is able to connect them: if not, the pair generates another tree (this applies recursively). If a resulting path is not found after a certain fixed number of iterations, the algorithm returns failure. Hauser, Brelt, Harada, Latombe - Using Motion Primitives in Probabilistic Sample-based Planning for Humanoid Robots.pdf

We will continue to look for other works. In the afternoon we will also write to professor Oriolo for some suggestions.

[matteosodano]

Motion Primitive Based Random Planning for Loco–Manipulation Tasks - Settimi, Caporale, Kryczka, Ferrati, Pallottino They introduce a mathematical background to partion a dynamical system in N coupled subsystems, each one subject to a different input. Each input is a loco-manipulation primitive that must keep the whole system stable. A solution is found by means of the P-Search algorithm, an extension of RRT. They also proved the probabilistic completeness and the asymptotic optimality of this algorithm. It checks feasibility and collision avoidance at each iteration. The code is available online. Settimi, Caporale, Kryczka, Ferrati, Pallottino - Motion Primitive Based Random Planning for Loco-Manipulation Tasks.pdf

[francesco-roscia]

Motion Planning of Ladder Climbing for Humanoid Robots - Zhang, Luo, Hauser, Ellenberg, Oh, Park, Paldhe, Lee Based on the approach in Hauser et al., they propose a motion primitive method for ladder climbing with humanoid robot. Primitives are defined for a single contact point. Chosen a primitive that connect two stances, collisions checking is performed and a feasible path for the joint variables is found by Inverse Kinematics. The main difference with the other papers is that each primitive has a nominal value but a randomly perturbed primitive is executed in order to find successful solutions. The perturbation is a random variable with uniform distribution on the interval [0, c], where c is chosen empirically. Zhang, Luo, Hauser, Ellenberg, Oh, Park, Paldhe, Lee - Motion Planning of Ladder Climbing for Humanoid Robots.pdf

Anytime Hybrid Driving-Stepping Locomotion Planning - Klamt, Behnke They divide the environment in cells and assign costs to the robot feet and base according to information on nearest cells. If some conditions hold, they allow the planner to consider steps for motion. They define abstract steps as the direct transition from a pre-step pose to a post-step pose. Each step introduces a certain cost. The result of the A* search is a cost-optimal abstract path which lacks executable motion sequences for steps and information about foot heights. The resulting path needs to be expanded during step motion generation. Stability and collision avoidance is enforced anytime. Klamt, Behnke - Anytime Hybrid Driving-Stepping Locomotion Planning.pdf

Other Approaches: Learning Based Techniques Many authors use different machine learning methods (clusterization, HMM, regression, NN) in biped locomotion. They can be useful in the definition of primitives (Dynamic Movement Primitives, DMPs), reconstructed from human movements. We did not go into details about this, because: 1) we cannot figure out how to collect data for centauro-like motions; 2) usually humanoids have less DOFs than humans, leading to unfeasible primitives for the robot. Should we go deeper?

Most of the articles we found propose similar (or equivalent) solutions to the ones of Oriolo et al., Hauser et al., and Klamt et al.