Technology

Five layers of neuromechanical intelligence.

We model 80 muscles with compliant Hill-type tendons — not rigid links. Each layer corresponds to a physiological level of human motor control.

05

Central Nervous System

Foundation Model — task-agnostic pretraining on GPU tendon simulation. 129D MSK observation → 80 coordinated muscle activations. 5.3B samples/day, zero labels. PMAAR‑TD shared subspace enables rapid fine-tuning for any task, user, or joint.

04

Excitation-Contraction Coupling

dA/dt = (S−A)/τ. Relaxation 4× slower than activation — a physiological constraint the controller must respect.

03

Hill-Type Compliant Tendon · Core Innovation

80 muscles × 9 parameters each. Fse = Fpe + Fce solved via fp64 Newton iteration on GPU. The only compliant tendon model running in GPU-parallel simulation.

02

Skeletal Dynamics

2,048 parallel worlds @ 600Hz physics. MuJoCo + NVIDIA Warp. 2.9× CPU throughput — 100× over single-threaded OpenSim.

01

Human-Exoskeleton Interface

πexo: 23D sensor → 2D motor current. End-to-end. INT8 quantized. Under 0.5ms embedded inference. One policy, no mode switching.

>10%
Metabolic Reduction
5.3B
Samples Per Day
<0.5ms
Inference Latency
9+
Modes · One Policy
A Different Approach

Beyond joint torque. Understanding muscle and tendon.

The exoskeleton industry treats the body as rigid links driven by joint torque — a convenient engineering simplification. But human movement isn't driven by joints. It's driven by 600+ muscles with compliant tendons that store and release elastic energy. When you ignore tendon dynamics, you fight the body's natural efficiency instead of augmenting it.

The Industry Standard

– Rigid-body joint-torque model — the body as a linkage
– Rule-based mode switching — preset torque curves, manual transitions
– Supervised learning — requires expensive labeled gait data from motion capture labs
– Joint-level control — each joint optimized independently, no muscle synergy
– Open-loop assistance — no learned intent recognition

Bones & Manifold

+ 80-muscle compliant Hill-type tendon model — elastic energy storage and release
+ Task-agnostic single neural network — 9+ modes, zero switching
+ RL + GPU simulation — 2,048 parallel worlds, zero labeling cost, 5.3B+ samples/day
+ Metabolic cost objective — directly optimizes human energy expenditure via RL
+ Learned intent recognition — the policy learns when to assist and when to wait
Products

Starting with the hip. Expanding to full lower limb.

Same AI core. Modular hardware. Consumer-first, not medical-first. Evolution roadmap →

First Batch
Hip
Outdoor
A second pair of legs for the trail
30Nm · 1.9kg · IP67 · Hot-swap battery
$1,199–1,999
First Batch
Hip
Urban
Invisible tech accessory
18Nm · 1.5kg · Wireless charging · Swappable shells
$699–1,299
First Batch
Hip
Active Senior
Let mom & dad climb that mountain again
20Nm · 1.8kg · Fall prevention · 6hr battery
$549–959
Second Batch
Hip+Knee
Care Senior
24/7 peace of mind
15Nm · SOS · 10hr battery · 4G · NMPA Class II
$549–825 or $41/mo
Second Batch
Spine
Youth
Scoliosis prevention, gamified
0.6kg · Flexible fabric · Posture gaming
$27–41
GenerationJointsMotorsSensorsComputeTarget
Gen 1HipAK60‑6 ×2BNO055 ×3 + Encoder ×2Teensy 4.12026
Gen 2Hip+KneeAK80‑9 ×4ICM‑20948 ×5 + EMGJetson Orin Nano2027
Gen 3Full Leg + AnkleAK80‑9 ×4 + Bowden ×2ICM‑20948 ×7 + EMG + GRFJetson Orin AGX2028
Scientific Foundations

Grounded in peer-reviewed research.

Nature 2024

Task-Agnostic Exoskeleton Control

(Molinaro et al.) Task-agnostic dynamics priors. 25 subjects, 28 activities, human+exoskeleton dataset. Foundation of our TCN baseline.

ICLR 2026

EXO-PLORE

(Song et al.) Human-aligned co-simulation framework. Validates that simulation-trained exoskeleton controllers transfer to real human experiments.

arXiv 2026

PMAAR-TD

(Wang et al.) Multi-agent average-reward TD-learning. 6-mode shared subspace. Personalized heads. Core RL framework.

Infrastructure

MyoAssist · MyoSuite

(Song et al.) Whole-body MSK modeling. 416 muscles with compliant Hill-type tendons. GPU-accelerated simulation upstream.

Foundation Model

Locomotion FM

(Wang & Son) Task-agnostic foundation model combining compliant tendon GPU simulation with PMAAR-TD shared subspace. 5.3B steps/day. Zero labels. Adapts to any task, user, or joint.

Validation

SMAT

(Song et al.) Staged human-exoskeleton co-adaptation. 10.1% hip muscle activation reduction. EMG validation.

Team

Engineers. Scientists. Founders.

CEO

Zirui Zhao

Columbia EE (ML). Studied under Prof. Shih-Fu Chang (NAE Member). Huawei AI Lab. 8yr CTO at Jiangsu Xinwang (formerly Shanghai Bell, Unitree Gold Partner) & Zhiqing Robot. 15yr algorithm engineering + hardware deployment.

CTO

Muxing Wang

NEU CS PhD Candidate. Edinburgh Statistics MSc (Distinction). PMAAR‑TD first author. Federated RL, multi-agent transfer learning. Shared subspace across 6 locomotion modes.

COO

Wenjing Yu

Peking Union Medical College. Shenyang Pharmaceutical University. Spring Sequence Capital Partner. MedTech investment & market strategy.

Academic Advisor

Seungmoon Song

NEU Assistant Professor. Move Lab PI. MyoAssist/MyoSuite creator. NIH K99/R00. Stanford postdoc. CMU PhD. Neuromechanical modeling.

Academic Advisor

Lili Su

NEU ECE Assistant Professor. Illinois PhD. MIT postdoc. NSF CAREER Award. Distributed ML, federated learning, multi-agent systems.

Academic Advisor

Pengkun Yang

Tsinghua Statistics Associate Professor. UIUC PhD. Princeton postdoc. High-dimensional statistics, meta-learning, convergence theory.

Columbia Northeastern Stanford Peking Union Medical College MIT Princeton Johns Hopkins Tsinghua Imperial College London Carnegie Mellon Illinois Urbana-Champaign Zhejiang University Huawei Microsoft Tesla NIO