Project Details
Projekt
Let AI choose its embodiment: Co-design of learning and mechatronics for fast reflexive robotics (CoLearnMech)
Applicant
Professor Dr. Sebastian Trimpe
Subject Area
Automation, Mechatronics, Control Systems, Intelligent Technical Systems, Robotics
Mechanics
Mechanics
Term
since 2025
Project identifier
Deutsche Forschungsgemeinschaft (DFG) - Project number 564414873
Imagine a robot hand holding a ball and flicking it with spin -- this exemplifies a robot task that requires fast and reflexive control. Similar tasks with instabilities, contact, and fast dynamics are common in promising robotics applications like industrial assembly, household services, or disaster response. While the high-level goal is usually clear (flick the ball into a bucket), breaking it down into a specific design is challenging because it involves both mechatronics (hardware) and control (software). Traditional approaches separate hardware and control design; this creates limitations and poses a "chicken and egg" problem: the simplest mechanical design may restrict sophisticated learning control, while constructing an overly complex anthropomorphic hand could be prohibitively expensive and impractical for effective learning. With this proposal, we strive for artificial intelligence (AI) methods that enable co-design of mechatronics and learning-based control in large design spaces. We formulate the co-design problem as optimizing a formalized objective across two levels, which we combine in our design assistant CoLearnMech. The inner level (AI1) learns to control a given design via learning-based control, in particular reinforcement learning (RL) and imitation learning (IL). In the outer level (AI2), Bayesian optimization (BO) proposes designs to AI1 and finds best designs based on achieved performance in AI1. The framework's ability to discover superior co-designs in high dimensional design spaces hinges on speedily solving AI1 and sample efficiency in AI2, which are two main gaps in existing methods that we address. To expedite solving AI1, we will develop data-efficient model-based RL methods, warm-start learning through IL demonstrations, and explore design-conditioned policies that generalize across multiple designs. As for AI2, we enhance sample efficiency through BO with partial information, branch-and-bound pruning of discrete variables, and critical feedback from human designers. To show the effectiveness of CoLearnMech, we push the limits of co-design in a real-world system: the design of a fast and reflexive robot hand, focusing on tasks like flicking a ball and manipulating objects in-hand. To validate that CoLearnMech can explore large design spaces and identify superior designs compared to existing commercial and open-source hands, we will conduct a large-scale benchmark simulation study and build hardware prototypes. In contrast to existing co-design methods that focus on detailed optimization of a few, continuous design variables, we aim for co-design at scale. We achieve this through a novel and modular co-design assistant, along with methodological contributions in RL, IL, and BO that enhance data efficiency, automation, and reliability as core challenges. With this approach, we intend to overcome the "chicken and egg" problem of complex robot design, as we will validate through the design of a fast robotic hand.
DFG Programme
Priority Programmes
