In robot manipulation, Reinforcement Learning (RL) often suffers from low sample efficiency and uncertain convergence, especially in large observation and action spaces. Foundation Models (FMs) offer an alternative, demonstrating promise in zero-shot and few-shot settings. However, they can be unreliable due to limited physical and spatial understanding. We introduce ExploRLLM, a method that combines the strengths of both paradigms. In our approach, FMs improve RL convergence by generating policy code and efficient representations, while a residual RL agent compensates for the FMs' limited physical understanding. We show that Explorllm outperforms both policies derived from FMs and RL baselines in table-top manipulation tasks. Additionally, real-world experiments show that the policies exhibit promising zero-shot sim-to-real transfer. Supplementary material is available at https://explorllm.github.io.

Individual pitch control (IPC) has been thoroughly researched for its ability to reduce wind turbine blade and tower fatigue loads. Conventional IPC often uses the multiblade coordinate (MBC) transformation and aims for full attenuation of the oscillating loads. However, this also leads to high control effort and increased fatigue damage on the pitch system. Output-constrained IPC uses the minimum actuator effort to drive loads to some reference value instead of fully attenuating them, achieving a trade-off between load reduction and actuator effort. To date, no control method exists that achieves output-constrained IPC using the conventional MBC approach. Furthermore, while multiple constrained IPC approaches have been proposed and analyzed, none of them analyze the full range of operating points between “no IPC” and “full IPC”. This paper presents two output-constrained IPC methods that use the MBC transformation. The first method, ℓ∞ IPC, independently drives the tilt and yaw moment to a tilt and yaw reference, while the second method, ℓ2 IPC, directly targets the magnitude of the combined tilt and yaw load. We furthermore analyze all operating points between no IPC and full IPC. OpenFAST simulations of the IEA 15 MW turbine were run at a wind speed of 15 m s−1. In laminar conditions, ℓ2 IPC is more efficient because it reduces the magnitude of the load directly, while ℓ∞ IPC also uses control effort to change the phase of the blade load in the direction of the load references. To assess the performance in realistic wind conditions, results are averaged over multiple turbulent wind seeds. Both ℓ∞ IPC and ℓ2 IPC have a similar performance, and the operating points between no IPC and full IPC form a nonlinear trade-off. One of the operating points in this trade-off achieves a 50 % load reduction, measured in damage equivalent load, with just 16.4 % of the actuator effort, measured in actuator duty cycle, compared to conventional IPC with the same integrator gain. This shows the potential of output-constrained IPC to facilitate a superior trade-off between load reduction and actuator effort.

Learning from demonstration is a promising approach for teaching robots new skills. However, a central challenge in the execution of acquired skills is the ability to recognize faults and prevent failures. This is essential because demonstrations typically cover only a limited set of scenarios and often only the successful ones. During task execution, unforeseen situations may arise, such as changes in the robot's environment or interaction with human operators. To recognize such situations, this paper focuses on teaching the robot situational awareness by using a camera input and labeling frames as safe or risky. We train a Gaussian Process (GP) regression model fed by a low-dimensional latent space representation of the input images. The model outputs a continuous risk score ranging from zero to one, quantifying the degree of risk at each timestep. This allows for pausing task execution in unsafe situations and directly adding new training data, labeled by the human user. Our experiments on a robotic manipulator show that the proposed method can reliably detect both known and novel faults using only a single example for each new fault. In contrast, a standard multi-layer perceptron (MLP) performs well only on faults it has encountered during training. Our method enables the next generation of cobots to be rapidly deployed with easy-to-set-up, vision-based risk assessment, proactively safeguarding humans and detecting misaligned parts or missing objects before failures occur.

This paper introduces a new imitation learning framework based on energy-based generative models capable of learning complex, physics-dependent, robot motion policies through state-only expert motion trajectories. Our algorithm, called Noise-conditioned Energy-based Annealed Rewards (NEAR), constructs several perturbed versions of the expert's motion data distribution and learns smooth, and well-defined representations of the data distribution's energy function using denoising score matching. We propose to use these learnt energy functions as reward functions to learn imitation policies via reinforcement learning. We also present a strategy to gradually switch between the learnt energy functions, ensuring that the learnt rewards are always well-defined in the manifold of policy-generated samples. We evaluate our algorithm on complex humanoid tasks such as locomotion and martial arts and compare it with state-only adversarial imitation learning algorithms like Adversarial Motion Priors (AMP). Our framework sidesteps the optimisation challenges of adversarial imitation learning techniques and produces results comparable to AMP in several quantitative metrics across multiple imitation settings. Code and videos available at anishhdiwan.github.io/noise-conditionedenergy-based-annealed-rewards/.

Objectives: To develop and validate a questionnaire on dental students' self-efficacy with tooth removal, suitable for measuring the effectiveness of training methods. Methods: To prepare and validate this questionnaire, we used the Association of Medical Education in Europe (AMEE) stepwise guide for developing questionnaires for educational research. In the validation process, our study group conducted two pilot studies, the first for an exploratory factor analysis and the second for a confirmatory factor analysis. In addition, the questionnaire was tested for convergence with the neuroticism subscale of the NEO-Personality Inventory. Results: After an exploratory factor analysis, which used a total of 137 responses on 33 items, 15 items were left for confirmatory factor analysis. A total of 118 responses were available for the confirmatory factor analysis. Model fitness was tested using tests for exact fitness and fit indices such as the goodness of fit index (GFI), root mean square error of approximation (RMSEA) and standardised root mean squared residual (SRMR). An acceptable fit was found for 11 items divided over three factors: ‘self-perceived skill’, ‘tension’ and ‘dedication’. These 11 items did not converge with the neuroticism scale. Conclusion: This study showed the development steps and initial validation of a psychometric instrument, the Amsterdam Self-Efficacy Scale for Tooth Removal (ASES-TR), consisting of 11 items for testing dental students' self-efficacy in performing tooth removal procedures.

Quadrupedal jumping has been intensively investigated in recent years. Still, realizing controlled jumping with soft landings remains an open challenge due to the complexity of the jump dynamics and the need to perform complex computations during the short time. This work tackles this challenge by leveraging trajectory optimization and behavior cloning. We generate an optimal jumping motion by utilizing dual-layered coarse-to-refine trajectory optimization. We combine this with a variable impedance control approach to achieve soft landing. Finally, we distill this computationally heavy jumping and landing policy into an efficient neural network via behavior cloning. Extensive simulation experiments demonstrate that, compared to classic model predictive control, the variable impedance control ensures compliance and reduces the stress on the motors during the landing phase. Furthermore, the neural network can reproduce jumping and landing behavior, achieving at least a 97.4% success rate. Hardware experiments confirm the findings, showcasing explosive jumping with soft landings and on-the-fly evaluation of the control actions.

Deep reinforcement learning (DRL) has emerged as a promising solution to mastering explosive and versatile quadrupedal jumping skills. However, current DRL-based frameworks usually rely on pre-existing reference trajectories obtained by capturing animal motions or transferring experience from existing controllers. This work aims to prove that learning dynamic jumping is possible without relying on imitating a reference trajectory by leveraging a curriculum design. Starting from a vertical in-place jump, we generalize the learned policy to forward and diagonal jumps and, finally, we learn to jump across obstacles. Conditioned on the desired landing location, orientation, and obstacle dimensions, the proposed approach yields a wide range of omnidirectional jumping motions in real-world experiments. In particular, we achieve a 90 cm forward jump, exceeding all previous records for similar robots. Additionally, the robot can reliably execute continuous jumping on soft grassy grounds, which is especially remarkable as such conditions were not included in the training stage.

This paper introduces a new imitation learning framework based on energy-based generative models capable of learning complex, physics-dependent, robot motion policies through state-only expert motion trajectories. Our algorithm, called Noise-conditioned Energy-based Annealed Rewards (NEAR), constructs several perturbed versions of the expert's motion data distribution and learns smooth, and well-defined representations of the data distribution's energy function using denoising score matching. We propose to use these learnt energy functions as reward functions to learn imitation policies via reinforcement learning. We also present a strategy to gradually switch between the learnt energy functions, ensuring that the learnt rewards are always well-defined in the manifold of policy-generated samples. We evaluate our algorithm on complex humanoid tasks such as locomotion and martial arts and compare it with state-only adversarial imitation learning algorithms like Adversarial Motion Priors (AMP). Our framework sidesteps the optimisation challenges of adversarial imitation learning techniques and produces results comparable to AMP in several quantitative metrics across multiple imitation settings.

Sim2real, the transfer of control policies from simulation to the real world, is crucial for efficiently solving robotic tasks without the risks associated with real-world learning. How-ever, discrepancies between simulated and real environments, especially due to unmodeled dynamics and latencies, significantly impact the performance of these transferred policies. In this paper, we address the challenges of sim2real transfer caused by latency and asynchronous dynamics in real-world robotic systems. Our approach involves developing a novel framework, REX (Robotic Environments with jaX), that uses a graph-based simulation model to incorporate latency effects while optimizing for parallelization on accelerator hard-ware. Our framework simulates the asynchronous, hierarchical nature of real-world systems, while simultaneously estimating system dynamics and delays from real-world data and implementing delay compensation strategies to minimize the sim2real gap. We validate our approach on two real-world systems, demonstrating its effectiveness in improving sim2real performance by accurately modeling both system dynamics and delays. Our results show that the proposed framework supports both accelerated simulation and real-time processing, making it valuable for robot learning.

Planning methods often struggle with computational intractability when solving task-level problems in large-scale environments. This work explores how the commonsense knowledge encoded in Large Language Models (LLMs) can be leveraged to enhance planning techniques for such complex scenarios. Specifically, we propose an approach that uses LLMs to efficiently prune irrelevant components from the planning problem's state space, thereby substantially reducing its complexity. We demonstrate the efficacy of our system through extensive experiments in a household simulation environment as well as real-world validation on a 7-DoF manipulator (video: https://youtu.be/6ro2UOtOQS4).

Sim2real, that is, the transfer of learned control policies from simulation to the real world, is an area of growing interest in robotics because of its potential to efficiently handle complex tasks. The sim2real approach faces challenges because of mismatches between simulation and reality. These discrepancies arise from inaccuracies in modeling physical phenomena and asynchronous control, among other factors. To this end, we introduce Engine Agnostic Graph Environments for Robotics (EAGERx), a framework with a unified software pipeline for both real and simulated robot learning. It can support various simulators and aids in integrating state, action, and time scale abstractions to facilitate learning. EAGERx’s integrated delay simulation, domain randomization features, and proposed synchronization algorithm contribute to narrowing the sim2real gap. We demonstrate (in the context of robot learning and beyond) the efficacy of EAGERx in accommodating diverse robotic systems and maintaining consistent simulation behavior. EAGERx is open source, and its code is available at https://eagerx.readthedocs.io

Several recent works show impressive results in mapping language-based human commands and image scene observations to direct robot executable policies (e.g., pick and place poses). However, these approaches do not consider the uncertainty of the trained policy and simply always execute actions that are suggested by the current policy as the most probable ones. This makes them vulnerable to domain shift and inefficient in the number of required demonstrations. We extend previous works and present the PARTNR algorithm that can detect ambiguities in the trained policy by analyzing multiple modes in the probability distributio of pick and place poses using topological analysis. In this way uncertainty in action can be estimated with single inference (and training single model) instead of using ensemble of models. Additionally, PARTNR employs an adaptive, sensitivity-based, gating function that decides if additional user demonstrations are required. User demonstrations are aggregated to the dataset and used for subsequent training. In this way, the policy can adapt promptly to domain shift and it can minimize the number of required demonstrations for a well-trained policy. The adaptive threshold enables to achieve the user-acceptable level of ambiguity to execute the policy autonomously and in turn, increase the trustworthiness of our system. We demonstrate the performance of PARTNR in a table-top pick and place task.

Learning from Interactive Demonstrations has revolutionized the way nonexpert humans teach robots. It is enough to kinesthetically move the robot around to teach pick-and-place, dressing, or cleaning policies. However, the main challenge is correctly generalizing to novel situations, e.g., different surfaces to clean or different arm postures to dress. This article proposes a novel task parameterization and generalization to transport the original robot policy, i.e., position, velocity, orientation, and stiffness. Unlike the state of the art, only a set of keypoints is tracked during the demonstration and the execution, e.g., a point cloud of the surface to clean. We then propose to fit a nonlinear transformation that would deform the space and then the original policy using the paired source and target point sets. The use of function approximators like Gaussian Processes allows us to generalize, or transport, the policy from every space location while estimating the uncertainty of the resulting policy due to the limited task keypoints and the reduced number of demonstrations. We compare the algorithm’s performance with state-of-the-art task parameterization alternatives and analyze the effect of different function approximators. We also validated the algorithm on robot manipulation tasks, i.e., different posture arm dressing, different location product reshelving, and different shape surface cleaning.

For an autonomous vehicle to operate reliably within real-world traffic scenarios, it is imperative to assess the repercussions of its prospective actions by anticipating the uncertain intentions exhibited by other participants in the traffic environment. Driven by the pronounced multi-modal nature of human driving behavior, this paper presents an approach that leverages Bayesian beliefs over the distribution of potential policies of other road users to construct a novel risk-aware probabilistic motion planning framework. In particular, we propose a novel contingency planner that outputs long-term contingent plans conditioned on multiple possible intents for other actors in the traffic scene. The Bayesian belief is incorporated into the optimization cost function to influence the behavior of the short-term plan based on the likelihood of other agents' policies. Furthermore, a probabilistic risk metric is employed to fine-tune the balance between efficiency and robustness. Through a series of closed-loop safety-critical simulated traffic scenarios shared with human-driven vehicles, we demonstrate the practical efficacy of our proposed approach that can handle multi-vehicle scenarios.

Developing physically assistive robots capable of dressing assistance has the potential to significantly improve the lives of the elderly and disabled population. However, most robotics dressing strategies considered a single robot only, which greatly limited the performance of the dressing assistance. In fact, healthcare professionals perform the task bimanually. Inspired by them, we propose a bimanual cooperative scheme for robotic dressing assistance. In the scheme, an interactive robot joins hands with the human thus supporting/guiding the human in the dressing process while the dressing robot performs the dressing task. We identify a key feature: the elbow angle that affects the dressing action and propose an optimal strategy for the interactive robot using the feature. A dressing coordinate based on the posture of the arm is defined to better encode the dressing policy. We validate the interactive dressing scheme with extensive experiments and also an ablation study.

With the aim of further enabling the exploitation of intentional impacts in robotic manipulation, a control framework is presented that directly tackles the challenges posed by tracking control of robotic manipulators that are tasked to perform nominally simultaneous impacts. This framework is an extension of the reference spreading (RS) control framework, in which overlapping ante- and post-impact references that are consistent with impact dynamics are defined. In this work, such a reference is constructed starting from a teleoperation-based approach. By using the corresponding ante- and post-impact control modes in the scope of a quadratic programming control approach, peaking of the velocity error and control inputs due to impacts is avoided while maintaining high tracking performance. With the inclusion of a novel interim mode, we aim to also avoid input peaks and steps when uncertainty in the environment causes a series of unplanned single impacts to occur rather than the planned simultaneous impact. This work in particular presents for the first time an experimental evaluation of RS control on a robotic setup, showcasing its robustness against uncertainty in the environment compared to three baseline control approaches.

Controlled execution of dynamic motions in quadrupedal robots, especially those with articulated soft bodies, presents a unique set of challenges that traditional methods struggle to address efficiently. In this study, we tackle these issues by relying on a simple yet effective two-stage learning framework to generate dynamic motions for quadrupedal robots. First, a gradient-free evolution strategy is employed to discover simply represented control policies, eliminating the need for a predefined reference motion. Then, we refine these policies using deep reinforcement learning. Our approach enables the acquisition of complex motions like pronking and back-flipping, effectively from scratch. Additionally, our method simplifies the traditionally labour-intensive task of reward shaping, boosting the efficiency of the learning process. Importantly, our framework proves particularly effective for articulated soft quadrupeds, whose inherent compliance and adaptability make them ideal for dynamic tasks but also introduce unique control challenges.

To promote the research in compliant quadrupedal locomotion, especially with parallel elasticity, we present Delft E-Go, which is an easily accessible quadruped that combines the Unitree Go1 with open-source mechanical add-ons and control architecture. Implementing this novel system required a combination of technical work and scientific innovation. First, a dedicated parallel spring with adjustable rest length is designed to strengthen each actuated joint. Then, a novel 3-D dual spring-loaded inverted pendulum model is proposed to characterize the compliant locomotion dynamics, decoupling the actuation with parallel compliance. Based on this template model, trajectory optimization is employed to generate optimal explosive motion without requiring reference defined in advance. To complete the system, a torque controller with anticipatory compensation is adopted for motion tracking. Extensive hardware experiments in multiple scenarios, such as trotting across uneven terrains, efficient walking, and explosive pronking, demonstrate the system’s reliability, energy benefits of parallel compliance, and enhanced locomotion capability. Particularly, we demonstrate for the first time the controlled pronking of a quadruped with asymmetric legs.

Formulating the dynamics of continuously deformable objects and other mechanical systems analytically from first principles is an exceedingly challenging task, often impractical in real-world scenarios. What makes this challenge even harder to solve is that, usually, the object has not been observed previously, and the only information that we can get from it is a stream of RGB camera data. In this study, we explore the use of deep learning techniques to solve this nonlinear identification problem. We specifically focus on extracting dynamic models of simple deformable objects from the high-dimensional sensor input coming from an RGB camera. We investigate a two-stage approach to achieve this goal. First, we train a variational autoencoder to extract an extremely low-dimensional representation of the object configuration. Then, we learn a dynamic model that predicts the evolution of these latent space variables. The proposed architecture can accurately predict the object's state up to one second into the future.

The advent of collaborative robots allows humans and robots to cooperate in a direct and physical way. While this leads to amazing new opportunities to create novel robotics applications, it is challenging to make the collaboration intuitive for the human. From a system’s perspective, understanding the human intentions seems to be one promising way to get there. However, human behavior exhibits large variations between individuals, such as for instance preferences or physical abilities. This paper presents a novel concept for simultaneously learning a model of the human intentions and preferences incrementally during collaboration with a robot. Starting out with a nominal model, the system acquires collaborative skills step-by-step within only very few trials. The concept is based on a combination of model-based reinforcement learning and inverse reinforcement learning, adapted to fit collaborations in which human and robot think and act independently. We test the method and compare it to two baselines: one that imitates the human and one that uses plain maximum entropy inverse reinforcement learning, both in simulation and in a user study with a Franka Emika Panda robot arm.