Human-Robot Integrated Simulation

KAIST Exoskeleton Lab

Human-Robot Integrated Simulation (HRIS)

Human Model for Human-Centered Physical AI

Physical AI systems that physically interact with humans must understand not only how humans move, but also how humans respond to robotic assistance. The Human-Robot Integrated Simulation (HRIS) team develops computational human models that interact dynamically with wearable robots in simulation. Instead of treating the human as a predefined motion trajectory, we model the human as an active agent whose motion emerges from neural motor control, musculoskeletal dynamics, sensory feedback, and physical interaction.

Closed loop between a human model and a wearable robot. The human controller sends muscle activation to the human plant and receives sensory feedback; the robot controller sends control input to the robot plant and receives the robot state; human plant and robot plant exchange interaction force.

Mechanism-Based Human Modeling

Many human simulations reproduce measured motion using motion capture data or reference trajectories. However, similar gait patterns may arise from different causes, such as muscle weakness, spasticity, joint contracture, or neurological impairment, and these differences can change the response to robotic assistance.

HRIS therefore focuses on modeling the mechanisms that generate motion, rather than simply reproducing motion outcomes. By incorporating pathological factors, motor-control mechanisms, musculoskeletal properties, and compensatory strategies, we aim to understand why a movement emerges and how it changes under robotic intervention.

Our research combines neuromusculoskeletal simulation, human motor control, biomechanics, and machine learning.

From gait imitation to mechanism-based human modeling. As-is: imitating only the measured gait trajectory leaves the pathological cause unmodeled and uninterpretable. To-be: a human model built from motor control theory and pathological factors, interacting with a wearable robot and external disturbances, makes the effect of the pathological cause predictable.

Modeling the Human from Neural Control to Physical Interaction

Our human model consists of two main components: the Human Controller and the Human Plant.

The Human Controller represents functional aspects of motor control, including voluntary motion generation, state estimation, balance control, disturbance response, and predictive correction. Rather than reproducing raw neural signals, we abstract functions associated with the motor cortical areas, basal ganglia, brainstem, cerebellum, and related systems into implementable control modules.

The Human Plant represents how motor commands are physically executed through muscle–tendon dynamics, passive joint and soft-tissue properties, posture-dependent moment arms, and spinal reflexes. Recent work includes a reduced neuromusculoskeletal model with passive muscle mechanics and stretch-reflex pathways implemented in NVIDIA Isaac.

Human model: the Human Controller (CNS) receives observations from the physics engine and sends muscle group activation to the Human Plant (musculoskeletal model), which returns its state and applies joint torque to the physics engine.
  • ControllerGenerates motor commands and corrective responses
  • PlantExecutes commands through body dynamics and environmental interaction

Toward Human-Centered Physical AI

The long-term goal of HRIS is to develop an integrated human–robot simulation framework that can represent interactions between wearable robots and humans with diverse physical and neurological characteristics. Through this framework, we aim to better understand individual differences in human motion and responses to robotic assistance, while providing a foundation for evaluating robot performance and safety across a wide range of human conditions before real-world deployment. Ultimately, our vision is to contribute to personalized wearable robotics and human-centered Physical AI that better account for human diversity.

Example 1

Neurological Functional Limitation: Parkinsonian Freezing of Gait Due to Basal Ganglia Dysfunction

Motion changes caused by manipulating the initiation gate (pathological functional parameter) of the basal ganglia model (motion regulator).

Human controller block diagram: sensory state estimation feeds the intent planner; its velocity command passes through the motion regulator (highlighted, the basal ganglia model) to the motor driver; error correction, posture stabilization and reflex control produce muscle activation for the human plant.

Real case

Freezing of gait

Basal-ganglia-limitation controller

Reduced step length, increased cadence
Example 2

Robot-Assistive Policy Learning for Right-Hemiplegic Stroke

No exo
Exo
Charts for a learned hip exoskeleton on frozen severe-stroke gait: hip flexion and extension without and with the exoskeleton, the applied exoskeleton hip torque, and paretic range of motion and symmetry.
+28%affected-side range of motion
45%of range-of-motion symmetry recovered
34%of stance-time symmetry recovered