Robot Control based on Motor Primitives:

One of the major challenges in robotics is generating complex motor behavior that matches human capabilities. Motor primitives are fundamental building blocks of a controller, enabling dynamic robot behavior with minimal high-level intervention. By treating motor primitives as basic “modules,” different modules can be sequenced or superimposed to generate a rich repertoire of motor behaviors.

Building Blocks of Neuromotor Control

The concept of motor primitives originates from human motor control research, where complex motor behavior in biological systems appears to be generated by combining fundamental building blocks. This idea dates back to Sherrington (1906), who proposed that reflexes are the basic units of motor behavior. Bernstein (1935) suggested that synergies serve as motor primitives to account for simultaneous joint movements or the activation of multiple muscles. More relevant to my research, Schaal (2004) and Hogan & Sternad (2007) proposed that rhythmic movements (e.g., for locomotion) and discrete movements (e.g., for reaching) are two distinct motor primitives. There is growing evidence that maintaining posture is distinct from controlling movement.

A significant advantage of motor primitives is that individual primitives can be combined to produce more complex behaviors. For example, Rohrer & Hogan (2004) and Flash (1991) have shown that the recovery of stroke survivors can be described by regaining the ability to superimpose multiple discrete motions to achieve a bell-shaped motion profile.

Modular Robot Control based on Motor Primitives

In our framework, called “Elementary Dynamic Actions” (EDA), we use discrete submovements, oscillations, and mechanical impedance as three distinct motor primitives. While oscillations and submovements describe the robot’s kinematics, mechanical impedance (stiffness, damping, and inertia) manages physical interaction. The key to this approach is breaking down complex physical interaction tasks into simpler, fundamental actions that can be more easily controlled and adapted. We combine the three primitives using an Equivalence Network model, which offers several benefits: 1) No need for inverse kinematics or dynamics; 2) No singularity handling; 3) Modularity!

More recently, we have demonstrated how the kinematic primitives of EDA can be efficiently combined with Contraction Theory and learned from demonstration by integrating them with the Dynamic Movement Primitives (DMP) framework.

Our recent work focuses on Dynamic Primitives (Impedance and Damping). By using peg-in-hole assembly, a common task in industrial assembly, we show that this assembly is feasible with a range of impedance parameter choices, challenging the conventional method of finding a single optimal setting. Moreover, we show that different peg types work for different assembly strategies and, more interestingly, the same assembly strategy works for all peg types! Our findings should support more streamlined and efficient programming practices in robotic assembly applications. For this, we introduce a neural network model as a success predictor for peg-in-hole assembly.

J. Lachner, F. Tessari, M.C. Nah, A.M. West, N. Hogan, “Divide et impera: Learning impedance families for peg-in-hole assembly,” submitted to T-RO, 2024. [LINK]

J. Lachner, M.C. Nah, F. Tessari and N. Hogan, “Elementary Dynamic Actions: key structures for contact-rich manipulation,” IROS Workshop CRM, 2023. [LINK]

M.C. Nah, J. Lachner, F. Tessari and N. Hogan, “Kinematic Modularity of Elementary Dynamic Actions,” IROS, 2024. [LINK]

M.C. Nah, J. Lachner and N. Hogan, “Robot control based on motor primitives: A comparison of two approaches,” The International Journal of Robotics Research, 2024. [LINK]

J. Lachner, F. Allmendinger, S. Stramigioli and N. Hogan, “Shaping Impedances to Comply With Constrained Task Dynamics,” in IEEE Transactions on Robotics, 2022. [LINK]