David W. Franklin

Human behaviour is developed through continuous adaptation to our environment over a range of timescales. Extensive studies have investigated the mechanisms and computations underlying this process of sensorimotor adaptation using several hundred trials. However, most of our motor skills have had countless hours of practice. Here we study a simple motor adaptation task using thousands of training trials over multiple weeks to study the long-term development of a motor memory, and examine changes in adaptation, retention, interlimb transfer, decay, spontaneous recovery and generalization. Unlike previous studies, participants showed near complete compensation to the novel dynamics, along with long-term increases in retention and spontaneous recovery. Moreover, we find narrowing in the angular generalization, suggesting continual tuning of the motor memory to the task. This demonstrates the extensive changes occurring with longer training of motor tasks, highlighting their importance in studies of sensorimotor control, rehabilitation and training.

Object handovers – while representing one of the simplest forms of physical interaction between two agents – involve a complex interplay of predictive and reactive control mechanisms in both agents. As human-human pairs have unrivaled skills in physical collaboration tasks, we take the approach of understanding and applying biomimetic concepts to human-robot interaction. Here, we apply the concept of passer movement cues, that is, slower movement for heavy objects and faster movements for lighter objects, to robot-human handovers. We first show that when a simulated passing agent's movement is scaled with object mass, participants as receivers adapt their anticipatory grip forces according to mass in a virtual environment. We then apply the same concept to a physical robot-human handover and show that our approach generalizes to the real-world. The predictive scaling of grip forces is learned iteratively upon repeated presentations of trajectory-mass pairings, whether the masses are presented in a random or blocked order. Overall we demonstrate that the presentation of robotic kinematic cues can provide intuitive and naturalistic human predictive control in object handover. This extends the use of non-verbal cues in robot-human handover tasks and facilitates more legible and efficient physical robot-human interactions.

: Humans often collaborate under asymmetric information, for example when two people carry a table and only one knows the destination. They coordinate without speech using cues from movement kinematics, interaction forces, and object states. Characterizing this sensorimotor communication is difficult because these signals both execute the task and convey information, whose meaning is context-dependent. Here, we investigated a virtual table-carrying task where one partner knew the target while the other inferred it from visuo-haptic feedback. Participants flexibly adapted kinematic and haptic cues across contexts to convey intention. We introduce an explainable machine-learning framework that decodes intent from ongoing multimodal signals and quantifies where individual features are informative. Incorporating the decoded signals into a drift-diffusion model accurately predicted the uninformed partner’s target choices and decision times. Together, our framework explains how humans communicate through action and offers principles for collaborative robots to infer and express intent through physical interaction.

Skilled control of visually-guided reaching is fundamental for many daily activities. Visual information about hand and target position are used for movement planning and online corrections through rapid visuomotor feedback responses. Such feedback control is generally believed to implicate a single error signal, representing a difference vector between hand and target position. Here, we directly assess whether shared or independent systems serve visually-guided feedback control. We tested whether feedback gains can be independently modulated by hand/cursor and target motion through manipulating the task-relevance of each signal during goal-directed reaching. Our results demonstrate that the gains of visuomotor feedback responses to perturbed hand and target motion can be set independently of one another, at the same time, as a function of task-relevance. By dissociating feedback control of cursor and target signals, our findings support the existence of two independent visuomotor feedback pathways, revealing a more flexible neural architecture for goal-directed action.

Motor skills such as bicycle riding are considered robust and transferable across bicycle types. However, when the steering direction is inverted (reversed bicycle) control is disrupted to the extent that the bicycle cannot be ridden. With sufficient practice, the reversed bicycle can be learned, but this learning appears to produce impairment of normal bicycle riding suggesting modification of this long-established motor memory. Here we investigate the learning process of riding a reversed bicycle over four days of practice, while repeatedly assessing normal bicycle performance to measure any potential interference. Introduction of the reversed bicycle disrupted predictive control, reflected in a consistently increased time lag in the steering-roll coupling during reversed bicycle trials. This increase in delay suggests that predictive behavior in normal bicycle riding cannot be transferred to the reversed bicycle. With training, some participants successfully learned to ride the reversed bicycle by gradually reorganizing this coupling, whereas others failed to acquire this inverted coupling. Notably, even short-term exposure to the reversed bicycle interfered with normal bicycle riding, reducing distance ridden and increasing variability in steering rate. Together, we show that even a highly practiced whole-body motor skill is susceptible to rapid interference when control dynamics are altered.

Muscle path modeling is more than just routing a cable that visually represents the muscle, but rather it defines how moment arms vary with different joint configurations. The muscle moment arm is the factor that translates muscle force into joint moment, and this property has an impact on the accuracy of musculoskeletal simulations. However, it is not easy to calibrate muscle paths based on a desired moment arm, because each path is configured by various parameters while the relations between moment arm and both the parameters and joint configuration are complicated. We tackle this challenge in the simple fashion of optimization, but with an emphasis on the gradient; when specified in its analytical form, optimization speed and accuracy are improved. We explain in detail how to differentiate the enormous cost function and how our optimization is configured, then we demonstrate the performance of this method by fast and accurate replication of muscle paths from a state-of-the-art shoulder-arm model. As long as the muscle is represented as a cable wrapping around obstacles, our method overcomes difficulties in path calibration, both for developing generic models and for customizing subject-specific models. This allows efficient enhancement of simulation accuracy for applications such as rehabilitation planning, surgical outcome prediction, and athletic performance analysis.

To precisely control our movements, the sensorimotor control system builds representations of the external world. However, the neural encoding of these representations – whether we encode our control as a function of specific kinematic or kinetic variables – and how we might then use these representations to build up complex actions is still unknown. Each prior study has provided support for a different reference frame or proposed mixture models to explain the complex findings. Here, we propose a different framework in which the sensorimotor control system uses multiple representations, allowing it to flexibly tune its representations to task demands. We directly test this by showing that the representation of external forces is totally governed by the nature of the experienced forces. Participants form a global force representation based on the coordinate system that best explains the forces experienced at different arm postures. We suggest that when formalizing such global representation, the system decides on the representation by considering the level of uncertainty it has in each of the coordinate systems. This flexibility and redundancy in the representations allows the precise tuning to any specific task within our complex environment, explaining our expertise in a wide range of tasks.

There is more to musculotendon path modeling than aligning a cable to reflect the geometric features of a musculotendon unit. From the perspective of simulation accuracy, the key is to replicate the length– and moment arm–joint angle relations of the target muscle. In this study, we propose an effect-oriented approach of automated path modeling, via hybrid calibration based on muscle surface mesh and moment arm. The task is formulated as a least-squares optimization problem with a threefold objective for the path to: (1) pass through multiple ellipses representing muscle cross-sections, (2) yield moment arms that match experimental measurements, and (3) yield moment arms in the expected functional directions. We demonstrate the performance of our optimization framework with the musculoskeletal surface mesh from the Visible Human Male and moment arm datasets from literature—producing 42 paths that are anatomically realistic and biomechanically accurate within 20.1 minutes. Our optimization framework is gradient-specified, which is faster and more accurate than using the default numerical gradient, making it applicable for large-scale subject-specific uses.

Humans have exceptional object manipulation skills. By combining feed-forward and feedback control, the sensori motor system is able to predictively scale grip and manipulation forces and quickly adapt to environmental changes. Using technologies such as virtual reality, researchers can investigate the underlying mechanisms in ways that are not possible in the physical world. Here, we present our custom-built virtual reality setup, including an open-source software framework, and show its validity in human motor control studies. We replicated two well-established experiments involving physical objects that investigated how humans adapt to different object mass and center of mass. Our results show that the general force and position control strategies employed in the virtual experiment closely mirrored those observed in the real world. Specifically, participants scaled grip forces with object mass and coordinated digit positions and forces according to the object's center of mass and shape. However, the trial-by-trial adaptation rate was slower, and the grip forces were slightly increased in the virtual setup, likely due to increased uncertainty arising from the 2D visual feedback and the lack of cutaneous feedback. Additionally, we tested the effect of visual feedback complexity by comparing a simple and detailed representation of the fingers, finding that participants exhibited similar manipulation strategies across both conditions. Our results validate this setup as a reliable platform for future studies in human motor control.

Haptic communication is a critical communication channel in physical collaboration. However, most studies focus on simplistic tasks with predetermined solutions, restricting the exploration of collaborative behaviors. In this study, we designed a complex task derived from the classic ball-beam control problem, requiring pairs of participants to collaboratively control an unstable object with internal degrees of freedom. The task’s redundant nature allowed diverse strategies and collaboration patterns to emerge. We systematically examined the impact of different sources of haptic feedback, distinguishing between forces arising from the collaborator’s actions and forces arising from the task dynamics. Participants collaborated across five haptic feedback conditions: full haptics, partner only, environment only, no haptics, or unrelated haptics. Haptic feedback significantly enhanced interpersonal coordination, primarily via partner-sourced rather than environment-sourced haptic feedback. Despite equal access to information and mechanical leverage, participants naturally assumed leader-follower roles, which were robustly quantified across multiple metrics. We showed how leadership dynamics evolved throughout the experiment and varied across different haptic conditions. This study advances the understanding of haptic communication, highlighting the value of partner haptics in developing more effective and adaptive strategies for both human-human and human-robot collaboration.