Research

Wearable robots that move with their users

Estimates indicate that by 2050 the U.S. will incur a two-fold increase in the incidence of amputation and stroke, due largely to the prevalence of vascular disease. These disabilities severely limit mobility and social activity for millions of Americans, whose ambulation is slower, less stable, and less efficient than that of able-bodied persons. The projected increase in mobility-related disabilities therefore presents a grand challenge to the American workforce and healthcare system. This motivates our research on high-performance wearable robots to enable mobility and improve quality of life for persons with disabilities.

Overview

Two paradigm shifts

Our work has driven two changes in how wearable robots are built and controlled. In prosthetics, control moved from discrete state machines that switch between pre-defined activities to continuous representations of phase and task. In exoskeletons, design moved from rigid, high-ratio actuators under trajectory-based control to backdrivable actuators under trajectory-free (task-agnostic) control. The funded projects below trace both.

A participant ascends laboratory stairs wearing the LocoLab Leg2 powered knee-ankle prosthesis.
Prosthetics: the LocoLab Leg2 on stairs under phase-based control
A person ascends laboratory stairs wearing a powered knee exoskeleton module.
Exoskeletons: a knee module on the same stairs under task-agnostic control
  1. 2013–2018

    Continuous phase, and the case for backdrivability

    NIH Director’s New Innovator Award · DP2HD080349 ProstheticsActuators

    Two foundations came out of one project. First, that progress through the gait cycle can be represented continuously by a phase variable derived from residual thigh motion, rather than by a state machine stepping through discrete events. Second, that increasing motor diameter while reducing gear ratio, at a fixed output torque, lowers actuator inertia and therefore backdrive torque. That second result is the design principle behind quasi-direct-drive actuation, which we translated to the wearable robotics field in 2017.

  2. 2017–2023

    Trajectory-free exoskeleton control

    NSF CAREER Award · CMMI-1652514 Exoskeletons

    Energy shaping alters the closed-loop dynamics of the body in an energetically passive manner, so the user retains control over energy growth and the device never imposes a joint trajectory. Body mass and inertia can be partially offloaded irrespective of the activity, reducing the activation of anti-gravity muscles across activities of daily life. The method was established through mathematical proofs and human-subjects experiments, but remained limited to specific joint configurations without biomimetic objectives.

  3. 2018–2023

    From continuous phase to a continuous task space

    NIH R01 · R01HD094772 Prosthetics

    Phase describes where in the stride the user is; task variables describe what the user is doing, including walking speed, ground or stair incline, and the transition between walking and stairs. Modeling joint patterns as continuous functions of both replaced a discrete set of activities with a continuum, and reduced the intent-recognition problem from five or more modes to three.

  4. 2021–2026

    Modular exoskeletons and task-agnostic assistance

    NIH R01 · R01EB031166 ExoskeletonsActuators

    A 480-gram quasi-direct-drive actuator mounted to conventional post-operative braces produced M-BLUE, a modular hip, knee, and ankle exoskeleton whose open-source designs have been downloaded more than 1,200 times. Data-driven energetic objectives let one controller assist across sit-stand transitions, level walking, ramps, and stairs without classifying the activity, followed by the first clinical trials in workers with quadriceps fatigue and in older adults.

  5. 2023–2028

    Impedance in place of kinematics

    NIH R01 renewal · R01HD094772 Prosthetics

    Strict kinematic control can produce abnormal forces between the prosthesis and the ground when timing, task estimates, or user posture carry small errors. Coupling joint impedance rather than joint kinematics to thigh motion lets the leg interact with the environment and adapt naturally under uncertainty, while kinematic control is retained during swing to preserve volitional placement of the foot.

  6. 2026–2030

    Powered unloader orthoses for osteoarthritis

    NIH R01 renewal · 2R01EB031166 Exoskeletons

    Extending backdrivable actuation and task-agnostic control to braces that actively offload arthritic joint surfaces. The work begins with knee osteoarthritis on the grounds of its prevalence, while advancing control technology intended to transfer to the hip and ankle.

01 · Prosthetics

Phase-based control of powered prosthetic legs

A powered prosthetic leg has motors that can actively push its user forward, but the assistance only helps if it arrives at the moment the user moves. Conventional controllers switch between a handful of preset behaviors, which compound across joints and tasks requiring hours of clinician tuning. Our controllers instead track the user's own residual thigh motion continuously, so the leg stays synchronized as the user speeds up, turns onto a ramp, or steps onto a staircase.

Three-panel diagram showing the progression from a discrete finite-state control paradigm, through a continuous phase variable derived from thigh angle, to a continuous task space mapped to joint kinematics.
Progression from a discrete to a continuous paradigm for modeling and control of human locomotion. Left: the conventional discretized paradigm, in which a high-level classifier selects an activity and a low-level state machine steps through phases of the gait cycle. Center: residual thigh motion converted into a continuous phase variable that repeats once per gait cycle. Right: a continuous task space related to the joint space over phase, shown here for walking at different ground inclines. The stairs variable represents steady-state walking at 0, steady-state stair climbing at 1, and transitions in between.

How it works

Finite state machines are typically used to switch controllers between discrete phases of the gait cycle, e.g., heel contact vs. toe contact, and between different tasks, e.g., uphill vs. downhill. However, this discrete methodology cannot continuously synchronize the prosthetic leg's motion to the timing or activity of the human user, and its complexity presents significant challenges with mode classification and patient-specific tuning.

We are investigating a continuous parameterization of human joint patterns based on 1) a phase variable that robustly represents the timing of the human gait cycle, and 2) task variables representing continuous variations of steady-state activities (e.g., walking at different speeds and inclines) or continuous transitions between activities (e.g., walk to stair ascent). Because the phase variable is driven by the user's own thigh, the user retains authority over the timing of every step, and the data-driven joint model supplies normative biomechanics without manual parameter tuning.

The same control stack has been deployed on four hardware platforms, including the second- and third-generation LocoLab legs, the Open-Source Leg, and the commercial Össur Power Knee. More than 20 research groups worldwide now use the phase-based paradigm, and we distribute our implementations as open-access libraries. We are also working with Össur to integrate the control stack into their next-generation products.

An above-knee amputee participant steps up onto a curb outdoors using the Össur Power Knee.
An above-knee amputee participant clearing a curb with the Össur Power Knee under phase-based control
Montage of four powered prosthetic legs: the LocoLab Leg2, Össur Power Knee, Open-Source Leg, and LocoLab Leg3 shown in a two-by-two grid at center, flanked by photographs of participants walking on each platform.
One control stack, four hardware platforms. Center: the LocoLab Leg2 (top left), the commercial Össur Power Knee (top right), the Open-Source Leg (bottom left), and the LocoLab Leg3 (bottom right). Of these two columns, the left holds powered knee-ankle prostheses and the right holds powered knee prostheses paired with passive ankle-foot components. The flanking photographs show participants walking on each platform under the same phase-based controller.
27%

lower peak hip flexion moment during swing while walking on the commercial Össur Power Knee under phase-based control, compared against the participants’ prescribed passive prostheses. Toe clearance rose by 19 mm and early-stance knee flexion by 6.3 degrees. Among the four higher-mobility participants, for whom the walking gains were concentrated, toe clearance rose by 25.4 mm and early-stance knee flexion by 7.1 degrees.

N=7 above-knee amputee participants · Best, Seelhoff, Wensman & Gregg, J. NeuroEngineering and Rehabilitation, 2025 · dataset released
23%

less time needed to rise from a seat than with the participants’ prescribed passive prostheses, with inter-limb loading asymmetry during standing falling by 30 percent, so weight was carried more evenly between the prosthetic and intact sides. Among the three lower-mobility participants, who rely on walking aids and stand to gain most, the time to rise fell by 36.7 percent.

N=7 above-knee amputee participants · Best, Seelhoff, Wensman & Gregg, J. NeuroEngineering and Rehabilitation, 2025 · dataset released

Key papers

JNER 2025 T-RO 2025 T-RO 2023 T-RO 2018 All 44 prosthetics papers

Videos

Capabilities of a novel phase-based control algorithm on the Össur Power Knee
Outdoor terrain challenges with the Össur Power Knee under phase-based control
The LocoLab Leg2 powered knee-ankle prosthesis across activities of daily life with above-knee amputee participants
Ambilateral activity recognition and continuous adaptation with a powered knee-ankle prosthesis — Cheng et al., IEEE Transactions on Robotics, 2025

This material is based upon work supported by the National Institutes of Health under Grant Numbers DP2HD080349 and R01HD094772 and by the National Science Foundation under Grant Numbers 1949346 and 2024237.

02 · Exoskeletons

Task-agnostic, energetic control of exoskeletons

Lower-limb exoskeletons, or powered orthoses, have potential to assist both unimpaired users in physically demanding jobs and impaired users navigating the home and community. An exoskeleton that plays back a recorded walking pattern works until its wearer does something else. Lifting a box from a low shelf, carrying it up a ramp, or standing from a chair requires a different assistance pattern, which is typically determined by a person pressing a button or a classifier that can make mistakes. Our controllers do not classify activity at all. They change the mechanics the wearer feels, reducing the effective weight of the body or the effective inertia of a limb, so one controller follows voluntary movement across ambulatory and non-ambulatory activities alike.

Diagram of a test circuit: a participant wearing a bilateral knee exoskeleton lifts a load, carries it over level ground, up stairs, across a raised platform, down a ramp, and lowers it again.
One trial, six tasks, no mode switching. A participant lifts a load (LL), carries it over level ground (LW), up the stairs (SA), down the ramp (RD), and lowers it again, with stair descent (SD) and ramp ascent (RA) on the return. The controller is never told which task is underway.
Left: hip, knee, and ankle exoskeleton modules worn over clothing, with the actuator, battery, and microcomputer labeled on each. Right: free-body diagrams contrasting open-loop body dynamics with closed-loop dynamics after energy shaping reduces apparent mass and gravity.
Shaping body energy through the Modular Backdrivable Lower-limb Unloading Exoskeleton (M-BLUE). Left: hip, knee, and ankle exoskeleton modules, each carrying its own actuator, battery, and microcomputer. Right: an example of energy-shaping control. Reducing gravity parameters in the potential energy reduces perceived body weight (solid force vectors), whereas reducing mass and inertia in the kinetic energy allows the user to accelerate limbs with less effort (dashed force vectors). Virtual nonlinear springs create variable extension torques during stance.

How it works

Traditional control methodologies for rehabilitation exoskeletons impose normative joint kinematic patterns on severely impaired users, and transitions between discrete tasks must be automatically classified, with inevitable errors, or manually triggered by the user or therapist. Emerging partial-assistance exoskeletons follow task-specific, time-based torque profiles that restrict users to a small set of pre-defined activities.

We are therefore investigating a task-agnostic control methodology that shapes the energetic dynamics of the human body with wearable actuators, using passivity theory to guarantee that the user remains in control of energy growth for stability. The paradigm can dynamically reduce mass and inertia parameters in body energetics, for example to offload the weight of a stroke patient during gait rehabilitation. Data-driven energy-shaping approaches provide a fraction of biological torque to reduce the muscle effort and joint loads that contribute to fatigue or arthritic pain.

Each controller is built by optimizing a combination of analytical basis functions, such as virtual nonlinear springs, dampers, and gravity and inertia compensation terms, to predict a scaled fraction of biological joint torque from limb kinematic and ground reaction force inputs over multi-activity human datasets. In its strictest form the exoskeleton can add net energy only by switching between stance and swing controllers, for example by changing the set point of a virtual spring, while relaxed formulations open designated power ports through which energy may be added continuously in a constrained manner. Relaxed passivity holds both within and outside the training data, enabling stability proofs for the coupled human-exoskeleton system and conferring robustness to the time delays and measurement errors introduced by compliant physical interfaces.

Powered ankle exoskeleton worn over a boot and shin guard.
Powered ankle exoskeleton
Ankle torque · one stride Exoskeleton assist
BIOLOGICAL WITH EXOSKELETON HEEL STRIKE PUSH-OFF SWING

Schematic, not measured data. The shaded band is the share of ankle torque the exoskeleton takes over, largest around push-off where the biological demand peaks. Across activities of daily life the measured reduction averaged 19.1 percent.

14.5%

reduction in quadriceps effort across lifting, lowering, and carrying tasks over level ground, ramps, and stairs, with mitigation of induced fatigue during repetitive lifting and lowering.

N=10 unimpaired adults · Divekar, Thomas, Yerva, Frame & Gregg, Science Robotics, 2024
19.1%

reduction in biological ankle torque across activities of daily life with the bilateral ankle exoskeleton.

N=10 unimpaired adults · Walters, Thomas & Gregg, IEEE Transactions on Robotics, 2026
24.7%

reduction in biological positive work at the hip, and 9.3 percent for the whole leg, while total biological plus exoskeleton hip power output increased to address diminished capacity in older adults.

N=8 older adults, NIH-funded clinical trial · preprint, manuscript under review

Key papers

Science Robotics 2024 TCST 2024 OJCSYS 2022 CSM 2018 All 33 exoskeleton papers

Videos

The knee exoskeleton assisting dynamic multi-terrain load carrying
The M-BLUE ankle, knee, and hip exoskeletons, as demonstrated at the Amazon MARS conference

This material is based upon work supported by the National Science Foundation under Grant Number 1652514 / 1949869 and by the National Institutes of Health under Grant Number R01EB031166.

03 · Actuators

High-torque, low-impedance wearable actuators

A motor in a wearable robot must be strong and light, which conventionally means a small fast motor behind a large gear reduction. That combination makes the joint feel rigid; the wearer cannot move it without the motor's cooperation. We take the opposite approach using high-torque motors with low-ratio transmissions, so the joint can be pushed by hand, swings freely, and absorbs impact, while still producing the torque a leg requires.

How it works

High-ratio transmissions result in high mechanical impedance, e.g., friction and reflected inertia, which means that the user cannot move, or backdrive, the joint without help from the actuator. Users with partial or full control of their legs require exoskeletons with backdrivable actuators to facilitate voluntary motion and comfort. Robotic prosthetic legs also need backdrivability to swing freely and to absorb forceful impacts for more natural amputee gaits.

Inspired by emerging legged robot designs, we design compact, lightweight, wearable actuators using custom high-torque motors with custom low-ratio transmissions, 24:1 or less, to achieve high output torques with very low mechanical impedance. Compared to conventional actuators, this class of quasi-direct-drive (QDD) actuators is quieter, more energy efficient, and better at controlling torque and impedance without a load cell. We have deployed these actuators on multiple generations of lower-limb exoskeletons and prosthetic legs, and we have released open-source designs for our modular hip, knee, and ankle exoskeleton, M-BLUE.

Quasi-direct-drive actuation is now becoming common in partial-assistance exoskeletons. We recently designed a powered knee prosthesis that weighs less than the Össur Power Knee while producing twice the torque, which demonstrates that the approach is commercially viable in prosthetics.

Computer-aided design view of the knee actuator assembly for the LocoLab Leg3 prosthesis.
Knee actuator assembly, LocoLab Leg3
Computer-aided design exploded view of a low-ratio compound gear train.
Low-ratio transmission, exploded CAD view

Key papers

RA-L 2022 T-MECH 2021 T-RO 2020 All 12 actuator papers Open-source CAD

Videos

A powered knee-ankle prosthesis with high-torque, low-impedance actuators — Elery et al., IEEE Transactions on Robotics, 2020
Assembling the modular, backdrivable hip and knee orthoses — Nesler et al., IEEE Robotics and Automation Letters, 2022

This material is based upon work supported by the National Institutes of Health under Grant Numbers DP2HD080349, R01HD094772, and R01EB031166.

04 · Ahead

Where this is going

Our foundational work in phase-based and task-agnostic control succeeded in the laboratory. The next problem is community-based utility, organized around three thrusts.

Continuously variable impedance control of prosthetic legs

A kinematic objective cannot respond naturally to forceful interactions with the user or the environment. We are modeling and controlling joint impedance, the relationship between joint motion and torque, over continuously varying locomotion, parameterizing stiffness, viscosity, and equilibrium angle as functions of phase and task variables such as speed, incline, and step height. In partnership with Össur we implement these impedance models on the Power Knee and evaluate them in clinical trials.

Generalizable task-agnostic control of exoskeletons

To scale task-agnostic assistance beyond the laboratory, a control framework must generalize across devices, joint configurations, and user populations. We are increasing the expressivity of energy shaping by replacing analytical energy basis functions with neural network representations, so the system can learn assistance patterns from multi-activity human data while mathematically maintaining the passivity guarantees required for safe physical human-robot interaction. The target is a device-agnostic control library that adapts to different sensing configurations and spans unimpaired through severely impaired users, accommodating a patient's recovery during rehabilitation.

Powered unloader orthoses for osteoarthritis

Current non-surgical management for osteoarthritis (OA) relies on passive orthoses that stabilize joints but fail to actively assist motion. We aim to establish a class of powered unloader braces that manage chronic pain by actively offloading joint articular surfaces, extending our backdrivable actuation and task-agnostic control to novel braces for both multi-compartmental knee osteoarthritis and ankle osteoarthritis. We hypothesize that offloading 15 to 30 percent of biological torque will significantly reduce painful joint loads, breaking the cycle of pain and inactivity and reducing the need for costly surgical intervention.