Soft microrobots based on photoresponsive materials and controlled by light fields can generate a variety of different gaits. This inherent flexibility be exploited to maximize their locomotion performance in given environment used adapt them changing conditions. Albeit, because the lack accurate models, intrinsic variability among microrobots, analytical control design is not possible. Common data-driven approaches, other hand, require running prohibitive numbers experiments lead very...

Changing conditions or environments can cause system dynamics to vary over time. To ensure optimal control performance, controllers should adapt these changes. When the underlying and time of change is unknown, we need rely on online data for this adaptation. In paper, will use time-varying Bayesian optimization (TVBO) tune in changing using appropriate prior knowledge objective its Two properties are characteristic many controller tuning problems: First, they exhibit incremental lasting...

Reinforcement learning (RL) aims to find an optimal policy by interaction with environment. Consequently, complex behavior requires a vast number of samples, which can be prohibitive in practice. Nevertheless, instead systematically reasoning and actively choosing informative gradients for local search are often obtained from random perturbations. These samples yield high variance estimates hence sub-optimal terms sample complexity. Actively selecting is at the core Bayesian optimization,...

When learning to ride a bike, child falls down number of times before achieving the first success. As falling usually has only mild consequences, it can be seen as tolerable failure in exchange for faster process, provides rich information about an undesired behavior. In context Bayesian optimization under unknown constraints (BOC), typical strategies safe explore conservatively and avoid failures by all means. On other side spectrum, non conservative BOC algorithms that allow failing may...

How can robots learn and adapt to new tasks situations with little data? Systematic exploration simulation are crucial tools for efficient robot learning. We present a novel black-box policy search algorithm focused on data-efficient improvements. The learns directly the treats as an additional information source speed up learning process. At core of algorithm, probabilistic model dependence parameters objective not only by performing experiments robot, but also leveraging data from...

Controller tuning is crucial for closed-loop performance but often involves manual adjustments. Although Bayesian optimization (BO) has been established as a data-efficient method automated tuning, applying it to large and high-dimensional search spaces remains challenging. We extend recently proposed local variant of BO include crash constraints, where the controller can only be successfully evaluated in an a-priori unknown feasible region. demonstrate efficiency through simulations...

Deep Reinforcement Learning (DRL) in simulation often results brittle and unrealistic learning outcomes. To push the agent towards more desirable solutions, prior information can be injected process through, for instance, reward shaping, expert data, or motion primitives. We propose an additional inductive bias robot learning: latent actions learned from demonstration as priors action space. show that these only a single open-loop gait cycle using simple autoencoder. Using combined with...

Diffusion models have recently gained popularity for policy learning in robotics due to their ability capture high-dimensional and multimodal distributions. However, diffusion policies are inherently stochastic typically trained offline, limiting handle unseen dynamic conditions where novel constraints not represented the training data must be satisfied. To overcome this limitation, we propose predictive control with (DPCC), an algorithm diffusion-based explicit state action that can deviate...

We consider the problem of sequentially optimizing a time-varying objective function using Bayesian optimization (TVBO). To cope with stale data arising from time variations, current approaches to TVBO require prior knowledge constant rate change. However, in practice, change is usually unknown. propose an event-triggered algorithm, ET-GP-UCB, that treats as static until it detects changes online and then resets dataset. This allows algorithm adapt realized temporal without need for...

Robust controllers ensure stability in feedback loops designed under uncertainty but at the cost of performance. Model time-invariant systems can be reduced by recently proposed learning-based methods, which improve performance robust using data. However, practice, many also exhibit form changes over time, e.g., due to weight shifts or wear and tear, leading decreased instability controller. We propose an event-triggered learning algorithm that decides when learn face LQR problem with rare...

We propose a data-driven control method for systems with aleatoric uncertainty, example, robot fleets variations between agents. Our leverages shared trajectory data to increase the robustness of designed controller and thus facilitate transfer new without need prior parameter uncertainty estimations. In contrast existing work on experience performance, our approach focuses uses collected from multiple realizations guarantee generalization unseen ones. is based scenario optimization combined...

Failures are challenging for learning to control physical systems since they risk damage, time-consuming resets, and often provide little gradient information. Adding safety constraints exploration typically requires a lot of prior knowledge domain expertise. We present measure which implicitly captures how the system dynamics relate set failure states. Not only can this be used as function, but also directly compute safe state-action pairs. Further, we show model-free approach learn by...

Changing conditions or environments can cause system dynamics to vary over time. To ensure optimal control performance, controllers should adapt these changes. When the underlying and time of change is unknown, we need rely on online data for this adaptation. In paper, will use time-varying Bayesian optimization (TVBO) tune in changing using appropriate prior knowledge objective its Two properties are characteristic many controller tuning problems: First, they exhibit incremental lasting...

Robust controllers ensure stability in feedback loops designed under uncertainty but at the cost of performance. Model time-invariant systems can be reduced by recently proposed learning-based methods, which improve performance robust using data. However, practice, many also exhibit form changes over time, e.g., due to weight shifts or wear and tear, leading decreased instability controller. We propose an event-triggered learning algorithm that decides when learn face LQR problem with rare...

Automated bin-picking is a prerequisite for fully automated manufacturing and warehouses. To successfully pick an item from unstructured bin the robot needs to first detect possible grasps objects, decide on object remove consequently plan execute feasible trajectory retrieve chosen object. Over last years significant progress has been made towards solving these problems. However, when multiple arms are cooperating decision planning problems become exponentially harder. We propose integrated...

Probabilistic models such as Gaussian processes (GPs) are powerful tools to learn unknown dynamical systems from data for subsequent use in control design. While learning-based has the potential yield superior performance demanding applications, robustness uncertainty remains an important challenge. Since Bayesian methods quantify of learning results, it is natural incorporate these uncertainties into a robust In contrast most state-of-the-art approaches that consider worst-case estimates,...