Three new robotics projects announced

Precision at a Distance (Lead: Colin Jones, Aude Billard)

Learning to Throw Under Uncertainty explores how robots can master the physics and adaptability of high-speed throwing—an ability that extends their workspace and enables agile, efficient interaction with distant or dynamic objects. The project brings together expertise in robotics and control to build systems that learn to throw accurately despite uncertainty in object properties, release timing, and motion. Using model-predictive and data-driven control, the research develops algorithms that combine physical modeling, real-time optimization, and experience-based learning to continually improve performance. The work culminates in a robotic-human juggling demonstration showcasing robust, adaptive coordination under uncertainty—a step toward machines that interact with the physical world with speed, precision, and reliability.

Advancing reinforcement learning and control for dexterous robotic manipulation (Lead: Maryam Kamgarpour, Josie Hughes)

This project focuses on advancing dexterous robotic manipulation using compliant robotic hands, targeting tasks such as object reorientation and precise grasping, with particular emphasis on long-horizon control and in hand manipulation. Reinforcement learning and imitation learning approaches such as behavioral cloning are widely used, but both suffer from inherent limitations when applied to complex manipulation problems. This motivates the development of hybrid methods that integrate the strengths of both frameworks: leveraging expert demonstrations to guide policies within the expert state distribution, while using reinforcement learning to fine-tune behavior and improve task success. The project will thus advance control and reinforcement learning theory and algorithms to address the challenges of dexterous, long-horizon robotic manipulation. The developed approaches will be verified on our in-house robotic hands.

Perception-aware Model Predictive Control with Guarantees (Lead: Melanie Zeilinger, Stephan Leutenegger) 

This project aims to develop a principled framework for safe and autonomous robotic exploration of unknown environments using vision-based sensing. The proposed approach leverages moving horizon estimation to robustly infer both the robot’s state and environmental features from visual-depth data. This estimation process is tightly integrated with a robust model predictive controller that ensures safety, relying solely on local state information without requiring global positioning. To guarantee complete exploration, active control strategies based on dual control principles are designed. The framework will be validated on both aerial and ground robotic platforms.