Embedded Files
How the Body Responds to Assistance
An assistive device does not simply add torque to a joint. When external forces act on the body, the nervous system adjusts how all the joints and muscles work together, and these adjustments can help or work against the device. Most devices are evaluated by averaging results across users. That average can hide the fact that each person may respond in their own way. We measure how people change their walking and muscle activity under assistance, at both the group level and the individual level, to understand what assistance really does to the body.
How we use it
Shared responses to hip assistance. We tested 22 healthy adults wearing a hip-assist soft exosuit, comparing walking without assistance to eight force profiles with different onset, peak and offset timing across the gait cycle. Almost every profile produced the same overall response. Peak hip flexion and ankle dorsiflexion increased, and soleus muscle activity decreased. The differences between profiles were small on average.
Individual signatures. When we looked at each stride, machine learning could hardly generalize responses from one person to another. Within each person, however, it could almost perfectly tell which force profile was being applied (macro-F1 > 0.98). Each person has a distinct, profile-dependent way of responding to assistance.
Adaptation over time. When people walked with one force profile for a longer period, the shared responses grew stronger. The body keeps adapting to assistance well after the first few steps.
What this means for device design. Responses to assistance have two layers: a shared response to the presence of assistance, and a personal response to how the assistance is timed. Assistive devices should therefore be evaluated and tuned for both. This finding motivates our work on personalized assistive control.
Simulation-Based Design of Assistive Control
Tuning an assistive device directly on people is slow and tiring. Every candidate controller needs repeated walking trials, and some settings may be unsafe. We first design and tune controllers on virtual humans. Our musculoskeletal models, driven by reinforcement-learning gait controllers, walk with a simulated device attached. We can then test thousands of controller settings and measure their effects on metabolic cost, joint loading and stability before anyone puts the device on. Only the most promising controllers move on to human experiments.
How we use it
Hip assistance without gait phase estimation. Most assistive controllers apply a preset torque profile timed to an estimated gait phase, which can fail when people turn or change their walking. We developed a modified delayed output feedback controller (MDOFC). It generates assistive torque from delayed and amplified hip joint angles, with added harmonic components so the torque profile can take more natural shapes. We optimized its two parameters in forward dynamics gait simulations. In simulation, MDOFC reduced metabolic cost more than the original controller and kept walking dynamically stable. In human experiments, it achieved a similar metabolic cost reduction with 8.8% less mechanical power.
Assisting impaired walking. We create patient models by weakening specific muscles, such as the hip muscles, and attach a multi-joint assistive device. A device controller is then trained with reinforcement learning to restore near-normal walking. This lets us design assistance for people with muscle weakness before clinical testing.
Personalized assistive control. People respond to assistance in their own ways. We are developing controllers that adapt to each person's response, combining simulation with experiments on wearable devices.
Soft Wearable Devices
Soft wearable devices, such as sleeves and suits made of flexible materials, can assist joints while staying light, comfortable and unobtrusive under clothing. Without a rigid frame, however, they cannot push on bones directly. Every force they produce must pass through skin and soft tissue before it reaches the joint. Their performance therefore depends on how the device conforms to and slides over the moving body. That interaction decides how much force becomes useful joint torque, how much pressure the wearer feels, and how strong the actuators must be. We develop simulation methods to design and evaluate soft wearable devices before any hardware is built.
How we use it
Simulating the body and the device together. We align a realistic body surface model (STAR) with a skeleton model so that the skin moves with the joints. We then fit a soft device around the limb and solve contact between its inner surface and the skin with an implicit contact method that remains stable as the device deforms. This gives assistive joint torque, skin pressure and the force each actuator must produce in one simulation.
A tessellated assistive sleeve for the knee. We proposed a tessellated hexahedral assistive sleeve (THAS). Thin hexahedral cells form a sleeve around the knee, and each cell can contract on its own. We compared contracting the front of the sleeve to assist knee extension with contracting the back to assist knee flexion.
Design guidance for actuators and comfort. The two strategies follow different mechanisms. Extension assistance is force-dominant: more torque requires stronger actuators and a careful distribution of pressure. Flexion assistance is stroke-dominant: more torque comes from a larger contraction range rather than more force. The predicted cell forces serve directly as actuator specifications. Peak pressures stayed below the discomfort threshold reported for healthy people.
A human model and its gait controller were created. A patient model was created by limiting the hip joint torque of the human model. A gait assistive device with four degrees of freedom was attached to the patient model. The controller of the device was trained using a reinforcement learning method to help the patient model walk normally. - by Jonghyun Park
Google Sites
Report abuse