The novelty-raahn algorithm has been shown to effectively learn a desired behavior from raw inputs by connecting an autoencoder with a Hebbian network. Hebbian learning is compelling for its biological plausibility and simplicity. It changes the weight of a connection based only on the activations of neurons it connects, and can effectively reinforce good behaviors when combined with neuromodulation. These low-level synaptic weight changes make for a better merge of the three learning tasks of perception, prediction and action. However, the state-ofthe art algorithm requires the design of a highly detailed modulation scheme designed for a specific system, which is disconnected from the overall objective it optimizes. In this thesis, we will propose that similar learning behavior can be achieved, by making the autonomous agent react to longer-term rewards, and thus implicitly introducing prediction capabilities. In doing so, the required modulation scheme becomes connected to the global optimization objective.

Current robots consume a lot of energy. The work required for a task is generally just a tiny fraction of the total energy consumed. Most energy wastefully dissipated. Design of the 2 degree of freedom (DoF) Plugless Robot Arm shows that clever design can reduce these energy losses to a tiny remaining fraction. The Plugless Arm is designed to perform a pick-and-place task displacing packages of 1 kg over a vertical distance of 1 m and some additional horizontal distance in a way that is capable of powering the full system. Maximum energy recovery is achieved by the design of springs in parallel to the actuators such that minimal actuator currents lead to minimal electric heat losses. This requires simultaneous optimisation of the trajectory and the parallel elastic elements. In a Cartesian configuration a novel method is tested which allows for implicit optimisation of an elastic element without additional parameterisation. The optimal parallel spring characteristic can be expressed as a pure function of the trajectory if the system dynamics are defined as a function of position instead of time. Afterwards the analytical optimal spring is replaced by a mechanism of low mechanical complexity showing similar energy characteristics. The mechanism is translated to a polar equivalent for the 2 DoF arm and optimised together with the trajectory of the arm. The optimal nominal trajectory is followed under undisturbed conditions when applying a pre-computed feedforward control signal to the actuators of the arm. Additional local optimal linear state feedback control is computed by means of Differential Dynamic Programming. All optimisation is done offline. Performance of the controller under disturbed conditions is tested in terms of accuracy as well as energy consumption. The system achieves a nominal energy retrieval of 3.64 J per cycle of 2.5 s, which is sufficient to power a light controller, the necessary sensors and a special energy efficient gripper. Some energy is left which can be used by the controller to recover from disturbances which extract energy from the system. The controller is also able to recover additional energy from disturbances which add energy to the system. The implementation of the controller is further optimised to achieve minimal computational energy consumption and memory requirements. By selective reduction of the control law, considerable data reduction is achieved with negligible impact on the controller performance.

Lyapunov's 2nd method can be formulated as a convex optimization problem by means of Sum-of-Squares (SOS) optimization. This gives a powerful tool for robust stability analysis of nonlinear systems. The use of SOS optimization is limited to small dynamical systems, because of the demanding numerical complexity underlying SOS optimization. For systems where scaling of SOS is a concern, diagonally dominant SOS (DSOS) is a possible alternative. DSOS approximates the SOS problem from within and results in a less complex optimization problem. The shortcoming of DSOS, however, is conservatism within the robust stability analysis, entailed by the approximation. In this thesis, we consider the problem of conservatism in robust stability analysis due to the application of DSOS. We propose new DSOS-based formulations which aim to reduce conservatism in robust stability analysis. Furthermore, we show how to formulate optimization problems which don't require an initial Lyapunov function candidate, and stabilize the numerical solutions of the DSOS problems. The main tool we leverage in this study is the Positivstellensatz in combination with DSOS. We demonstrate the proposed formulations on three numerical experiments. We estimate the region of attraction of the van der Pol oscillator, investigate stability of an uncertain nonlinear system and proof stability of a nonlinear system, which is documented to suffer from numerical issues when approached with SOS methods. The results show a reduction in conservatism and improvement in numerical robustness when compared to the DSOS formulation found in literature.

Nowadays, the number of robotic systems grows enormously. In order to execute their task, many of these robots need a fixturing gripper to grasp and hold objects. To reduce the environmental impact, the operational costs, the size of actuators and to improve the uptime of mobile robots, it becomes increasingly important to design robots for a minimum energy consumption. Therefore this thesis is dedicated to developing an energy saving robotic gripper. The gripper's drive has to enable the gripper to perform sizing}, the adjustment to the object size, and forcing, exerting retaining forces on the object. As drive strategy an actuator in combination with an internal force mechanism was selected, because this allows for autonomous grasping and passive retention in any orientation. Based on a newly proposed performance metric, which normalizes the product of the minimum dislodging force in any direction and the average closing speed with the rated actuator power, a Statically Balanced Force Amplifier (SBFA) was selected as internal force mechanism. In addition to current state-of-the-art drives, which are able to switch between two transmission branches for either high-speed or high-force and overpower the motor because of static-load cancellation with dedicated locking mechanisms, an SBFA realizes a force reduction of approximately 95%. A concept for a drive is developed based on this strategy. By analyzing the three subfunctions of the drive: 1) sizing, 2) forcing and 3) switching between sizing and forcing, five design principles are derived. The SBFA is analyzed under the non-ideal conditions as occur in a gripper drive: a finite object stiffness and a position error in the size adjustment. A topology selection is conducted, which resulted in a concept for a drive consisting of an SBFA with a movable frame and a differential that resolves the motor input in a forcing and sizing branch. The timing of the switching is controlled by an additional non-linear threshold spring and an endstop is implemented to bypass the main spring during switching. An actual prototype for the implementation in the Delft Hand 3 is designed and realized. The mechanism has a cylindrical housing with a diameter of 64 mm and a length of 78.5 mm and weights approximately 460 grams. Experiments are conducted which validate the working principle, i.e. the ability to perform sizing, forcing and switching, of the developed prototype. The maximum output torque is 1.073 Nm. For 5 emulated object sizes, an average torque reduction of 91.5 % is measured. This prototype validates the feasibility to combine an SBFA, NBDM and two different transmission branches. Future work is required to further improve the functionality and asses the performance of the drive.

Kinodynamic planning is motion planning in state space and aims to satisfy kinematic and dynamic constraints. To reduce its computational cost, a popular approach is to use sampling based methods such as RRT with off-line machine learning for estimating the steering cost and inputs. However, scalability and robustness are still open challenges in these type of Learning-RRT algorithms. We propose the use of generative adversarial networks (GAN) for learning of the steering cost and inputs. Furthermore, a novel data generation method is introduced, which is easy to learn and, in terms of parameter count, scales linearly to higher degrees of freedom. In our experiments, we show that the GAN has excellent generalisation capabilities, resulting in a considerable improvement in performance compared to the state-of-the-art. Consequently, we show that our method can scale to a planar arm and is robust to data dimensionality.

Kinodynamic motion planning for a robot involves generating a trajectory from a given robot state to goal state while satisfying kinematic and dynamic constraints. Rapidly-exploring Random Trees (RRT) is a sampling-based algorithm that has been widely adopted for this. However, RRT is not fast enough to enable its use in industrial applications. Recently, supervised learning has been used to pre-learn time consuming steps of RRT which resulted in improvement in planning times. The supervised learning models require cost and control input of the system as training data which are generated using optimal control. The training data can be obtained either by indirect optimal control or direct optimal control techniques. In this thesis, both the techniques are each used to generate cost and control inputs for a two-link manipulator using random initial-final state pairs. Then each dataset is used to train a model and the datasets are compared based on certain training metrics. K-nearest neighbours regression and multi-layer perceptron neural network are the supervised learning models used in this thesis. It is observed that both the datasets result in similar convergence of the models, but indirect optimal control approach allows upto 24-fold faster data generation and upto 3-fold reduction in dimensionality of training data compared to the direct optimal approach. Real-world robots have torque limits based on actuator configuration. The torque limits are modeled as control constraints in both the optimal control techniques and the effect of this restriction on data generation and supervised learning is studied in this thesis. Direct optimal control is found to be better for data generation in this case due to the ease of applying control bounds as inequality constraints on the function approximations. Indirect optimal control is very tedious as active constraints should be known a priori to determine the switching points. An alternate method is explored instead where samples are generated similar to the unconstrained case but samples violating the constraints are removed. Poor control input learning is observed in both approaches and the models struggled to extrapolate. It is hypothesised that this is due to inability of the constrained data to fully capture the system dynamics. However, good cost prediction is achieved using neural networks.