In this paper, a hierarchical and robust framework for learning bipedal locomotion is presented successfully implemented on the 3D biped robot Digit built by Agility Robotics. We propose cascade-structure controller that combines process with intuitive feedback regulations. This design allows to realize stable walking reduced-dimensional state action spaces of policy, significantly simplifying increasing sampling efficiency method. The inclusion regulation into improves robustness learned...

This paper presents a novel model-free reinforcement learning (RL) framework to design feedback control policies for 3D bipedal walking. Existing RL algorithms are often trained in an end-to-end manner or rely on prior knowledge of some reference joint trajectories. Different from these studies, we propose policy structure that appropriately incorporates physical insights gained the hybrid nature walking dynamics and well-established zero approach As result, overall has several key...

This paper presents a novel reinforcement learning (RL) framework to design cascade feedback control policies for 3D bipedal locomotion. Existing RL algorithms are often trained in an end-to-end manner or rely on prior knowledge of some reference joint task space trajectories. Unlike these studies, we propose policy structure that decouples the locomotion problem into two modules incorporate physical insights from nature walking dynamics and well-established Hybrid Zero Dynamics approach...

In this work, we demonstrate robust walking in the bipedal robot Digit on uneven terrains by just learning a single linear policy. particular, propose new control pipeline, wherein high-level trajectory modulator shapes end-foot ellipsoidal trajectories, and low-level gait controller regulates torso ankle orientation. The foot-trajectory uses policy regulator PD law. As opposed to neural network based policies, proposed has only 13 learnable parameters, thereby not guaranteeing sample...

The design of feedback controllers for bipedal robots is challenging due to the hybrid nature its dynamics and complexity imposed by high-dimensional models. In this paper, we present a novel approach using Reinforcement Learning (RL) Hybrid Zero Dynamics (HZD). Existing RL approaches walking are inefficient as they do not consider underlying physics, often requires substantial training, resulting controller may be applicable real robots. HZD powerful tool control with local stability...

This paper studies the class of scenario-based safety testing algorithms in black-box configuration. For sharing same state-action set coverage with different sampling distributions, it is commonly believed that prioritizing exploration high-risk state-actions leads to a better efficiency. Our proposal disputes above intuition by introducing an impossibility theorem provably shows all aforementioned difference perform equally well expected Moreover, for covering sets state-actions,...

This work presents a hierarchical framework for bipedal locomotion that combines Reinforcement Learning (RL)-based high-level (HL) planner policy the online generation of task space commands with model-based low-level (LL) controller to track desired trajectories. Different from traditional end-to-end learning approaches, our HL takes insights angular momentum-based linear inverted pendulum (ALIP) carefully design observation and action spaces Markov Decision Process (MDP). simple yet...

This project describes the design and construction of an Automatic Voltage Regulator (AVR) Electronic Load Controller (ELC) for voltage frequency regulation in island Micro-hydropower Plant (MHP). For control, speed by ballast load method has been used. To this approach, a combined binary-continuous was employed. The implemented AVR is totally self-excited means energy transfer system which allows isolated operation MHP. entire designed considering current standard regulations Ecuadorian...

In this paper, with a view toward deployment of light-weight control frameworks for bipedal walking robots, we realize end-foot trajectories that are shaped by single linear feedback policy. We learn policy via model-free and gradient free learning algorithm, Augmented Random Search (ARS), in the two robot platforms Rabbit Digit. Our contributions two-fold: a) By using torso support plane orientation as inputs, achieve robust on slopes upto 20° simulation. b) demonstrate additional behaviors...

The dynamic response of the legged robot locomotion is non-Lipschitz and can be stochastic due to environmental uncertainties. To test, validate, characterize safety performance robots, existing solutions on observed inferred risk incomplete sampling inefficient. Some formal verification methods suffer from model precision other surrogate assumptions. In this paper, we propose a scenario based testing framework that characterizes overall by specifying (i) where (in terms set states)...

This paper presents a neural-network based adaptive feedback control structure to regulate the velocity of 3D bipedal robots under dynamics uncertainties. Existing Hybrid Zero Dynamics (HZD)-based controllers through implementation heuristic regulators that do not consider model and environmental uncertainties, which may significantly affect tracking performance controllers. In this paper, we address uncertainties in robot from perspective reduced dimensional representation virtual...

Safe path planning is critical for bipedal robots to operate in safety-critical environments. Common algorithms, such as RRT or RRT*, typically use geometric kinematic collision check algorithms ensure collision-free paths toward the target position. However, approaches may generate non-smooth that do not comply with dynamics constraints of walking robots. It has been shown control barrier function (CBF) can be integrated RRT/RRT*to synthesize dynamically feasible paths. Yet, existing work...

The design of feedback controllers for bipedal robots is challenging due to the hybrid nature its dynamics and complexity imposed by high-dimensional models. In this paper, we present a novel approach using Reinforcement Learning (RL) Hybrid Zero Dynamics (HZD). Existing RL approaches walking are inefficient as they do not consider underlying physics, often requires substantial training, resulting controller may be applicable real robots. HZD powerful tool control with local stability...

Safe navigation in real-time is an essential task for humanoid robots real-world deployment. Since are inherently underactuated thanks to unilateral ground contacts, a path considered safe if it obstacle-free and respects the robot's physical limitations underlying dynamics. Existing approaches often decouple planning from gait control due significant computational challenge caused by full-order robot In this work, we develop unified, framework that can be evaluated online real-time,...

Linear policies are the simplest class of that can achieve stable bipedal walking behaviors in both simulation and hardware. However, a significant challenge deploying them widely is difficulty extending to more dynamic like hopping running. Therefore, this work, we propose new linear which template models be embedded. In particular, show how embed Spring Loaded Inverted Pendulum (SLIP) model policy realize perpetual arbitrary directions. The spring constant learned addition remaining...

Dynamic locomotion in legged robots is close to industrial collaboration, but a lack of standardized testing obstructs commercialization. The issues are not merely political, theoretical, or algorithmic also physical, indicating limited studies and comprehension regarding standard infrastructure equipment. For decades, the approaches we have been were rarely standardizable with hand-pushing, foot-kicking, rope-dragging, stick-poking, ball-swinging. This paper aims bridge gap by proposing use...

This work presents a hierarchical framework for bipedal locomotion that combines Reinforcement Learning (RL)-based high-level (HL) planner policy the online generation of task space commands with model-based low-level (LL) controller to track desired trajectories. Different from traditional end-to-end learning approaches, our HL takes insights angular momentum-based linear inverted pendulum (ALIP) carefully design observation and action spaces Markov Decision Process (MDP). simple yet...

This paper presents a novel framework for learning robust bipedal walking by combining data-driven state representation with Reinforcement Learning (RL) based locomotion policy. The utilizes an autoencoder to learn low-dimensional latent space that captures the complex dynamics of from existing data. reduced dimensional is then used as states training RL-based gait policy, eliminating need heuristic selections or use template models planning. results demonstrate learned variables are...

This paper presents a novel model-free reinforcement learning (RL) framework to design feedback control policies for 3D bipedal walking. Existing RL algorithms are often trained in an end-to-end manner or rely on prior knowledge of some reference joint trajectories. Different from these studies, we propose policy structure that appropriately incorporates physical insights gained the hybrid nature walking dynamics and well-established zero approach As result, overall has several key...

This paper presents a neural-network based adaptive feedback control structure to regulate the velocity of 3D bipedal robots under dynamics uncertainties. Existing Hybrid Zero Dynamics (HZD)-based controllers through implementation heuristic regulators that do not consider model and environmental uncertainties, which may significantly affect tracking performance controllers. In this paper, we address uncertainties in robot from perspective reduced dimensional representation virtual...

The dynamic response of the legged robot locomotion is non-Lipschitz and can be stochastic due to environmental uncertainties. To test, validate, characterize safety performance robots, existing solutions on observed inferred risk incomplete sampling inefficient. Some formal verification methods suffer from model precision other surrogate assumptions. In this paper, we propose a scenario based testing framework that characterizes overall by specifying (i) where (in terms set states)...