Learning Complex Motor Skills for Legged Robot Fall Recovery

Article Computer Science, Software Engineering

ASE: Large-Scale Reusable Adversarial Skill Embeddings for Physically Simulated Characters

Xue Bin Peng et al.

Summary: The study presents a large-scale data-driven framework for learning versatile and reusable skill embeddings for physically simulated characters. The models can be trained using unstructured motion clips without any task-specific annotation or segmentation, and can learn a rich and versatile repertoire of skills. The system allows users to specify tasks through simple reward functions, and the skill embedding enables the character to automatically synthesize complex and naturalistic strategies to achieve the task objectives.

ACM TRANSACTIONS ON GRAPHICS (2022)

Add to Collection

Article Robotics

Trajectory Optimization with Optimization-Based Dynamics

Taylor A. Howell et al.

Summary: The article presents a framework for bi-level trajectory optimization, where a system's dynamics are encoded as a solution to a constrained optimization problem and the smooth gradients of this problem are passed to an upper-level trajectory optimizer. This approach allows for constraint handling, additional variables, and non-smooth behavior to be abstracted from the upper-level optimizer and enables classical unconstrained optimizers to synthesize trajectories for complex systems.

IEEE ROBOTICS AND AUTOMATION LETTERS (2022)

Add to Collection

Proceedings Paper Automation & Control Systems

Accessibility-Based Clustering for Efficient Learning of Locomotion Skills

Chong Zhang et al.

Summary: The study focuses on algorithmic improvements in data efficiency and robustness simultaneously by automatically discovering initial states using the K-Access algorithm. Results show faster learning and high success rate in fall recovery compared to baselines.

2022 IEEE INTERNATIONAL CONFERENCE ON ROBOTICS AND AUTOMATION (ICRA 2022) (2022)

Add to Collection

Proceedings Paper Automation & Control Systems

Sim-to-Real Learning of All Common Bipedal Gaits via Periodic Reward Composition

Jonah Siekmann et al.

Summary: By defining a parameterized reward function using simple probabilistic periodic costs, we successfully achieved the transfer of various gaits of bipedal robots from simulation to reality. This method can assist in learning different gaits and is applicable to common bipedal locomotion.

2021 IEEE INTERNATIONAL CONFERENCE ON ROBOTICS AND AUTOMATION (ICRA 2021) (2021)

Add to Collection

Proceedings Paper Automation & Control Systems

DeepWalk: Omnidirectional Bipedal Gait by Deep Reinforcement Learning

Diego Rodriguez et al.

Summary: This paper presents a novel DRL approach for enabling robots to learn omnidirectional locomotion for humanoid robots by gradually increasing task difficulty. The method does not require reference motions, reducing overall complexity, and provides strategies for sim-to-real transfer.

2021 IEEE INTERNATIONAL CONFERENCE ON ROBOTICS AND AUTOMATION (ICRA 2021) (2021)

Add to Collection

Article Robotics