International Journal of Humanoid Robotics

Humanoid robots are required to perform a wide repertoire of tasks working beside humans in complex dynamic environments. Learning mechanisms are important for building up these types of repertoires of robot skills. However, despite the clear advantages of these approaches, it would be impractical to teach the robot skills for every needed task and for every foreseen situation. Robot skills learning approaches to develop humanoid robotic systems would have greater impact if the models of the skills can be operated upon to generate new behaviors of increasing levels of complexity. A framework that allows the adaptation of previously learned motion skills to new unseen contexts is necessary. In this work, we present different modalities for the adaptation and generation of new skill models based on the already learned models of the skills. Experimental results are presented to validate this approach.

Most robots are today controlled as being entirely rigid. But often, as for HRP-2 robot, there are flexible parts, intended for example to absorb impacts. The deformation of this flexibility modifies the orientation of the robot and endangers balance. Nevertheless, robots have usually inertial sensors inertial measurement units (IMUs) to reconstruct their orientation based on gravity and inertial effects. Moreover, humanoids have usually to ensure a firm contact with the ground, which provides reliable information on surrounding environment. We show in this study how important it is to take into account these information to improve IMU-based position/orientation reconstruction. We use an extended Kalman filter to rebuild the deformation, making the fusion between IMU and contact information, and without making any assumption on the dynamics of the flexibility. We show how, with this simple setting, we are able to compensate for perturbations and to stabilize the end-effector's position/orientation in the world reference frame. We show also that this estimation is reliable enough to enable a closed-loop stabilization of the flexibility and control of the center of mass (CoM) position with the simplest possible model.

Recent developments in human–robot interaction show how the ability to communicate with people in a natural way is of great importance for artificial agents. The implementation of facial expressions has been found to significantly increase the interaction capabilities of humanoid robots. For speech, displaying a correct articulation with sound is mandatory to avoid audiovisual illusions like the McGurk effect (leading to comprehension errors) as well as to enhance the intelligibility in noisy conditions. This work describes the design, construction and testing of an animatronic talking face developed for the iCub robot. This talking head has an articulated jaw and four independent lip movements actuated by five motors. It is covered by a specially designed elastic tissue cover whose hemlines at the lips are attached to the motors via connecting linkages. The mechanical design and the control scheme have been evaluated by speech intelligibility in noise (SPIN) perceptual tests that demonstrate an absolute 10% intelligibility gain provided by the jaw and lip movements over the audio-only display.

This paper presents a framework for dynamic walking on uneven terrain using a novel time-varying extension of the divergent component of motion (DCM). By varying the natural frequency of the DCM, we are able to achieve generic CoM height trajectories during stepping. The proposed approach computes admissible DCM reference trajectories given desired zero moment point (ZMP) plans for single and double support, permitting both flat-footed and heel-toe walking. Real-time planning is accomplished using reverse-time integration of the discretized DCM dynamics over a finite time horizon. To account for discontinuities during replanning, linear model predictive control (MPC) is implemented over a short preview window, enabling smooth transitions between steps. DCM tracking control is achieved using a time-varying proportional-integral controller based on the virtual repellent point (VRP). The effectiveness of the combined approach is verified in simulation using a 30 DOF model of THOR, a compliant torque-controlled humanoid. We demonstrate dynamic locomotion on uneven terrain and heel-toe walking using a complementary whole-body controller to track the corresponding VRP forces.

In this paper, we propose an integrated policy learning framework that fuses iterative learning control (ILC) and reinforcement learning. Integration is accomplished at the exploration level of the reinforcement learning algorithm. The proposed algorithm combines fast convergence properties of iterative learning control and robustness of reinforcement learning. This way, the advantages of both approaches are retained while overcoming their respective limitations. The proposed approach was verified in simulation and in real robot experiments on three challenging motion optimization problems.

Robotic learning on real hardware requires an efficient algorithm which minimizes the number of trials needed to learn an optimal policy. Prolonged use of hardware causes wear and tear on the system and demands more attention from an operator. To this end, we present a novel black-box optimization algorithm, Reward Optimization with Compact Kernels and fast natural gradient regression (ROCK^⋆). Our algorithm immediately updates knowledge after a single trial and is able to extrapolate in a controlled manner. These features make fast and safe learning on real hardware possible. The performance of our method is evaluated with standard benchmark functions that are commonly used to test optimization algorithms. We also present three different robotic optimization examples using ROCK^⋆. The first robotic example is on a simulated robot arm, the second is on a real articulated legged system, and the third is on a simulated quadruped robot with 12 actuated joints. ROCK^⋆ outperforms the current state-of-the-art algorithms in all tasks sometimes even by an order of magnitude.

We explore how to represent, plan and learn robot pouring. This is a case study of a complex task that has many variations and involves manipulating non-rigid materials such as liquids and granular substances. Variations of pouring we consider are the type of pouring (such as pouring into a glass or spreading a sauce on an object), material, container shapes, initial poses of containers and target amounts. The robot learns to select appropriate behaviors from a library of skills, such as tipping, shaking and tapping, to pour a range of materials from a variety of containers. The robot also learns to select behavioral parameters. Planning methods are used to adapt skills for some variations such as initial poses of containers. We show using simulation and experiments on a PR2 robot that our pouring behavior model is able to plan and learn to handle a wide variety of pouring tasks. This case study is a step towards enabling humanoid robots to perform tasks of daily living.

Humanoid robots that have to operate in cluttered and unstructured environments, such as man-made and natural disaster scenarios, require sophisticated sensorimotor capabilities. A crucial prerequisite for the successful execution of whole-body locomotion and manipulation tasks in such environments is the perception of the environment and the extraction of associated environmental affordances, i.e., the action possibilities of the robot in the environment. We believe that such a coupling between perception and action could be a key to substantially increase the flexibility of humanoid robots.

In this paper, we approach the affordance-based generation of whole-body actions for stable locomotion and manipulation. We incorporate a rule-based system to assign affordance hypotheses to visually perceived environmental primitives in the scene. These hypotheses are then filtered using extended reachability maps that carry stability information, for identifying reachable affordance hypotheses. We then formulate the hypotheses in terms of a constrained inverse kinematics problem in order to find whole-body configurations that utilize a chosen set of hypotheses.

The proposed methods are implemented and tested in simulated environments based on RGB-D scans as well as on a real robotic platform.