Bility Targets Amputee Fall Risks With Bio-Robotic Locomotion AI

The Architecture of Motion in Motion: How to Harness Physical AI for Next-Era Human-Aware — Ai4

Ai4

Assistive bio-robotics requires a dedicated physical AI foundation to anticipate human intent within tens of milliseconds, because unlike autonomous humanoids with full body control, prosthetic balance failures cause catastrophic injuries rather than software bloopers.

AI-assisted summary of The Architecture of Motion in Motion: How to Harness Physical AI for Next-Era Human-Aware — original source. Check the original for full context.

Source published: 2026-08-15

Detailed notes

## Background Published by Ai4 on YouTube on August 15, 2026, this presentation features Yishay Falick, Co-Founder and CEO of Bility. The session explores how data-driven physical AI and bio-robotics can bridge the gap between human biology and machine intelligence to solve the global mobility crisis and restore human dignity for amputees.

## TL;DR Yishay Falick challenges the prevailing 'completeness fallacy' in embodied AI, contrasting self-contained autonomous humanoid robots with the asymmetric partnership required when assistive AI shares a physical body with a human. Operating under strict sub-100-millisecond physiological deadlines, bio-robotic prostheses must anticipate intent, accommodate compensated gaits, and prevent catastrophic falls. To resolve the lack of standardized locomotion datasets, Bility combines real-world empirical capture from diverse patient populations with NVIDIA-accelerated physics simulations to generate petabytes of edge-case training data. Falick presents Bility's vision of an open platform consisting of a foundational human locomotion model, a body-aware operating system, and advanced metamaterial hardware designed to serve as shared infrastructure across all co-embodied AI applications.

## Atmosphere & style The presentation carries an urgent, sobering, and deeply compassionate tone that cuts sharply through typical robotics hype. By contrasting the triviality of a humanoid robot's lab blooper reel with the devastating human cost of an elderly amputee falling on a street curb, Falick grounds complex biomechanical engineering in visceral human stakes. The talk feels intellectually rigorous yet emotionally stirring, balancing technical discussions of control theory and federated edge sensing with a compelling, mission-driven appeal for ecosystem-wide collaboration.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=49VC0ofQaO4&t=4s) Human-assistive physical AI operates under radically higher stakes than autonomous robotics: while a humanoid falling in a lab is logged as a software bug or blooper reel, an amputee falling risks severe injury or permanent loss of independence. - [00:04](https://www.youtube.com/watch?v=49VC0ofQaO4&t=4s) Bility operates under the foundational hierarchy of 'data first, model second, hardware third,' combining ground-truth lab captures with physics-constrained simulation. - [01:35](https://www.youtube.com/watch?v=49VC0ofQaO4&t=95s) Autonomous humanoids rely on a 'completeness assumption' where a unified model owns every actuator, whereas bio-robotic prosthetics must solve an asymmetric 'partnership problem' controlling a single limb attached to an independent biological body. - [01:35](https://www.youtube.com/watch?v=49VC0ofQaO4&t=95s) Locomotion operates on strict physiological deadlines: edge AI must sense, infer, and actuate within tens of milliseconds before the biological nervous system's ~100ms error-detection latency threshold. - [03:24](https://www.youtube.com/watch?v=49VC0ofQaO4&t=204s) The mobility crisis is predominantly driven by chronic health conditions rather than trauma: diabetes causes 50% of all amputations, and lower-limb loss accounts for 75% of total cases. - [03:24](https://www.youtube.com/watch?v=49VC0ofQaO4&t=204s) Current prosthetics perpetuate a 'dignity gap' by using rigid, sensor-blind sockets that fail to adapt to fluctuating biological tissue volumes or variable terrains. - [05:03](https://www.youtube.com/watch?v=49VC0ofQaO4&t=303s) Bility's AI architecture integrates environmental terrain sensing, muscle intent tracking (EMG), and real-time personalization engines to query a foundational human locomotion model for multi-joint torque actuation. - [06:42](https://www.youtube.com/watch?v=49VC0ofQaO4&t=402s) Unlike computer vision (ImageNet) or NLP, human locomotion historically lacked a standardized foundation dataset encompassing diverse body types, compensated pathologies, and fatigue states. - [06:42](https://www.youtube.com/watch?v=49VC0ofQaO4&t=402s) To safely train models on dangerous fall scenarios, Bility pairs empirical mobility lab data with an NVIDIA Inception-powered physics simulator to generate petabytes of synthetically validated edge cases. - [09:07](https://www.youtube.com/watch?v=49VC0ofQaO4&t=547s) A bio-robotic operating system treating body awareness as a first-class primitive is being developed alongside advanced metamaterials and energy-harvesting structural components. - [10:53](https://www.youtube.com/watch?v=49VC0ofQaO4&t=653s) The locomotion corpus and operating system serve as general embodied AI infrastructure across co-embodied applications, including industrial exoskeletons and rehabilitation robotics. - [13:46](https://www.youtube.com/watch?v=49VC0ofQaO4&t=826s) Bio-robotic devices employ federated sensor grids combining EMG intent arrays and pressure matrices, providing immediate out-of-the-box balance while continuously learning the user's specific gait patterns. - [16:03](https://www.youtube.com/watch?v=49VC0ofQaO4&t=963s) Lower limbs are prioritized over upper limbs because legs bear full body weight, represent 75% of amputation cases, and carry immediate life-or-death consequences from falls. - [17:06](https://www.youtube.com/watch?v=49VC0ofQaO4&t=1026s) Bility positions itself as an open platform enabler, commercializing its locomotion datasets, operating system, and bio-robotic simulator to foster an industry-wide ecosystem of adaptive mobility hardware.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=49VC0ofQaO4&t=4s) The Stakes of Physical AI: Human Dignity vs. The Blooper Reel

Yishay Falick introduces a category of artificial intelligence that the mainstream technology sector has rarely encountered. It is neither a large language model predicting the next token nor an autonomous vehicle pathfinder navigating city traffic. Instead, it is physical AI engineered to share a physical body with a living human being, requiring millisecond-level precision on every single step.

The distinction between autonomous humanoid robotics and human-assistive physical AI comes down to consequences and body ownership. When a bipedal humanoid robot loses balance and falls inside an engineering lab, it is treated as a software bug, logged into a bug tracker, and compiled into an entertaining blooper reel. When a 68-year-old man stumbles while stepping off a street curb, the outcome is often a fractured hip, emergency hospital admission, and the permanent loss of personal independence. While both scenarios share identical Newtonian physics and control theory challenges, the stakes are completely different because an autonomous robot owns and coordinates its entire mechanical body, whereas a human-assistive device controls only an isolated fraction of an external biological organism.

Bility approaches this challenge not simply as a traditional prosthetics manufacturer, but as a physical AI data company. Founded on the principle of "data first, model second, hardware third," the company operates a specialized mobility laboratory capturing ground-truth kinematic data from healthy subjects, military veterans, and diabetic amputees. To scale this empirical data, Bility pairs lab captures with a physics-constrained simulator capable of generating petabytes of edge-case locomotion scenarios that cannot be gathered manually. Bility has been part of the NVIDIA Inception program to build out this data-driven foundation.

- **Foundational Hierarchy**: Data first, model second, hardware third. - **The Core Platform**: Empirical motion capture lab paired with an NVIDIA Inception-accelerated physics simulator.

---

[01:35](https://www.youtube.com/watch?v=49VC0ofQaO4&t=95s) The Completeness Fallacy and the Partnership Problem

The prevailing paradigm in embodied AI rests on robot foundation models: end-to-end neural policies mapping sensory perception directly to motor torque output. This paradigm succeeds due to a quiet, unstated assumption: completeness. In a standard robot, the model owns every actuator, calculates its exact center of mass across all internal linkages, and maintains unified intent across its entire mechanical frame.

```mermaid flowchart TD accTitle: Robotics Completeness vs Bio-Robotic Partnership subgraph Humanoid["Autonomous Robot: The Completeness Paradigm"] A["Unified AI Model"] --> B["Full Body Actuation"] B --> C["Total Center of Mass Control"] end subgraph BioRobotic["Bio-Robotic Prosthetic: The Partnership Problem"] D["Prosthetic AI Policy"] --> E["Single Actuated Limb"] E -. "Blind to Spine / Hip / Shoulder".-> F["Biological Human Body"] F -. "EMG & Intent Feedback".-> D end ```

Deploying an embodied AI policy onto an amputee completely invalidates the completeness assumption. The model controls only a single replacement limb while remaining blind to the biological hip, spine, and shoulders it cannot directly actuate. Locomotion ceases to be a closed-loop control problem and becomes an asymmetric "partnership problem." The machine must continuously infer human intent in advance, execute anticipatory actions, tolerate prediction errors, and recover before biological sensory feedback loops detect an anomaly.

Human locomotion operates on strict physiological deadlines: - **Stride Duration**: A typical human walking stride lasts approximately 1.0 second. - **Biological Latency**: The human nervous system requires roughly 100 milliseconds just to register that a misstep or mechanical fault has occurred. - **Edge Inference Window**: The onboard AI must complete the entire cycle—sense, infer, decide, and actuate—within tens of milliseconds on every step. - **Operating Environment**: The system must sustain this cycle over 10,000 steps per day across 20 years, functioning reliably in driving rain, over loose gravel, and at 3:00 AM, with "no cloud in that loop, no human in that loop, no second take."

---

[03:24](https://www.youtube.com/watch?v=49VC0ofQaO4&t=204s) The Reality of Limb Loss: Diabetes, Rigidity, and the Dignity Gap

The scale of the mobility crisis is driven largely by demographic and chronic health shifts. Approximately 2.25 million Americans live with limb loss today, a figure projected to climb to 3.5 million by 2050. Lower-limb amputations account for three out of every four cases. Public perception frequently associates limb loss exclusively with military combat trauma; while veterans represent a critical population, combat and physical trauma account for only one in eight amputations. Diabetes is the primary driver, causing fully half of all amputations.

Regardless of whether an amputee is a 24-year-old soldier recovering from a blast injury or a 68-year-old patient who lost a foot to a non-healing diabetic ulcer, their core requirement is identical: total, unhesitating trust in the prosthetic limb.

``` +-------------------------------------------------------------------------------+ | THE CURRENT PROSTHETIC REALITY | +-------------------------------------------------------------------------------+ | Rigid Socket Fit Cast once; fails as biological tissue swells/shrinks | | Static Configurations Single stiffness setting for office, stairs, beach | | Sensor Blindness Zero diagnostic feedback on diabetic tissue health | | Aesthetic Disconnect Unnatural appearance drives devices into closets | +-------------------------------------------------------------------------------+ ```

The current standard of care forces users to strap rigid composite sockets onto dynamic, living tissue that naturally swells and contracts throughout the day. This static mechanical interface causes chronic pain, soft tissue breakdown, and high abandonment rates. Wearers are given a single rigid mechanical configuration to navigate vastly different environments—office carpets, staircases, sandy beaches, and dance floors. Devices remain heavy, expensive, and sensor-blind, failing to monitor the underlying pathology of the limb. This failure represents a fundamental "dignity gap" rather than just a hardware limitation.

---

[05:03](https://www.youtube.com/watch?v=49VC0ofQaO4&t=303s) Conceptual Architecture: The Locomotion Model Pipeline

To eliminate the dignity gap, Bility designed a patent-pending conceptual architecture that replaces passive mechanical hardware with an active, human-aware robotic system.

```mermaid flowchart LR accTitle: Bility Conceptual AI Locomotion Pipeline subgraph Inputs["Perception & Awareness"] Env["Environmental Sensing\n(Terrain & obstacles)"] Human["Human Awareness\n(Kinematics & EMG intent)"] end

subgraph Processing["Inference & Personalization"] Pred["Prediction &\nPersonalization Engine"] Loco["Trained Human\nLocomotion Model"] end

subgraph Output["Actuation"] Act["Millisecond Multi-Joint\nTorque Actuation"] end

Env --> Pred Human --> Pred Pred --> Loco Loco --> Act ```

The physical architecture incorporates dynamic robotic interfaces instead of rigid caps, autonomous motorized joints, environmental terrain sensors, integrated diagnostic sensing to monitor tissue health, a shape-shifting foot, and programmable aesthetic visibility.

On the computational side, environmental perception maps upcoming obstacles and surfaces while human-awareness sensors track the user's kinematic trajectory and muscle intent. These streams feed into an inference engine driving predictive gait personalization, adapting continuously because "your gait is not my gait and your gait tonight is not your gait this morning." This engine queries the underlying human locomotion foundation model to deliver sub-millisecond motor actuation. While drawing the architectural diagram is simple, training the bottom-level locomotion foundation model represents the true engineering bottleneck of embodied AI.

---

[06:42](https://www.youtube.com/watch?v=49VC0ofQaO4&t=402s) Solving the Data Bottleneck: Ground Truth and Physics Simulation

Transformative breakthroughs across artificial intelligence have historically occurred only when comprehensive, standardized datasets were assembled: - **Computer Vision**: Unlocked by ImageNet. - **Natural Language Processing**: Unlocked by the open internet. - **Protein Folding**: Unlocked by 50 years of structural crystallography data.

Human locomotion has lacked any comparable foundation: there has been no unified, labeled dataset recording how human limbs articulate across diverse body types, variable terrains, muscular fatigue states, and compensated pathologies.

```mermaid flowchart TD accTitle: Locomotion Data Engine Architecture subgraph Capture["Empirical Capture (Mobility Lab)"] H["Healthy Subject Baselines\n(Natural gait mechanics)"] C["Compensated Gait Capture\n(Amputees, Veterans, Diabetics)"] E["Distribution Hot Spots\n(Curbs, wet tile, micro-stumbles)"] end

subgraph Synthesis["Physics-Constrained Simulation"] Engine["NVIDIA Inception Simulation Stack"] Rules["Strict Physics Constraints:\nMass, Inertia, Ground Reaction, Energy Cost"] end

subgraph Output["Unified Foundation Corpus"] Data["Petabytes of Verified Training Data"] end

H --> Engine C --> Engine E --> Engine Engine --- Rules --> Data ```

Bility structured its business around creating this locomotion corpus. In its mobility lab, the company captures baseline kinematic data from healthy subjects alongside real-world data from amputees, veterans, and vasculitis/diabetic patients who guard painful socket interfaces or alter their mechanics due to exhaustion. Training models on textbook biomechanics alone fails because a prosthetic must accommodate the user's actual compensated gait rather than an idealized theoretical stride.

The dataset deliberately prioritizes the distributional "hot spots" where falls occur: misjudged curbs, slick flooring, and recovered micro-stumbles. Because vulnerable patients cannot be subjected to thousands of physical laboratory falls, Bility feeds empirical captures into a physics-constrained simulation engine running on the NVIDIA Inception stack. The engine generates petabytes of synthetic edge cases under strict physical validation—enforcing realistic mass distributions, rotational inertia, ground reaction forces, and biological metabolic costs: "Real data teaches the model what is true. Simulation teaches it what is possible."

---

[09:07](https://www.youtube.com/watch?v=49VC0ofQaO4&t=547s) A Body-Aware Operating System and Material Engineering

A bio-robotic system is not a standalone neural network; it is an integrated network composed of an actuated knee, an ankle, an active socket interface, real-time sensor grids, a high-efficiency power bus, and an autonomous human user. Bility is developing a dedicated operating system for bio-robotics that treats body awareness as a first-class primitive, featuring a real-time data grid and built-in physics simulation. The platform mirrors what ROS achieved for autonomous mobile robotics, but incorporates a dynamic node that ROS never had to account for: "a living person with opinions."

Parallel hardware engineering tracks support the platform across two domains: - **Advanced Metamaterials**: Structural architectures engineered to minimize device weight while maintaining high fatigue strength, improving dynamic stability, and offering programmable aesthetic visibility so materials actively perform functional tasks rather than merely enclosing internal components. - **Prototype Lower-Leg Engineering**: A full-stack lower-limb prosthesis combining synthetic structural bone, sensor-embedded dermal layers, and an energy-harvesting power supply that recharges onboard reserves directly from the mechanical ground pressure generated during walking. Bility manufactures this complete physical device to validate the motion platform under real-world human loads.

---

[10:53](https://www.youtube.com/watch?v=49VC0ofQaO4&t=653s) Embodied AI Infrastructure: Walking With Humans, Not For Them

The mechanical and computational frameworks developed for bio-robotics address the broader challenge of co-embodied AI: systems where autonomous machines share physical loads directly with human anatomy. This includes industrial exoskeletons, physical rehabilitation robotics, powered tactical suits, and active assistive wearables. A unified human locomotion corpus and a body-aware operating system represent shared embodied AI infrastructure, with prosthetics serving as the most latency-critical and high-consequence use case.

```mermaid flowchart TD accTitle: Broad Embodied AI Infrastructure Reach Core["Bility Physical AI Core\n(Human Locomotion Corpus + Body-Aware OS)"] Core --> P["Lower-Limb Prosthetics\n(Highest consequence & latency)"] Core --> E["Industrial & Medical Exoskeletons"] Core --> R["Rehabilitation Robotics"] Core --> W["Active Assistive Wearables & Hard Suits"] ```

Falick emphasizes that data hoarding in human mobility is unacceptable when millions of individuals face daily mobility limitations. While robotics investments focus heavily on autonomous machines moving through environments alone, millions of humans require robotic systems that walk *with* them:

> "Millions of people will never need a robot that walks for them. They need one that walks with them. That is the harder problem. You do not own the whole body. You do not get a second take. And being wrong is not a blooper reel. It is a person on the floor."

The impact of this approach is illustrated by a real-world scenario: a woman steps off a curb that drops 9 inches rather than the 6 inches her vision anticipated. Within tens of milliseconds, the prosthetic AI identifies the spatial error, dynamically recalculates joint stiffness, recirculates mechanical energy, and centers her mass over the foot. The user continues walking forward without breaking stride or ever realizing that a fall was prevented, restoring dignity, fluidity, and athletic grace.

---

[13:46](https://www.youtube.com/watch?v=49VC0ofQaO4&t=826s) Q&A: Federated Sensor Orchestration vs. Monolithic Humanoids

During the transition to the audience Q&A session, Falick elaborates on the structural differences between autonomous bipedal humanoids and human-interactive bio-robotics:

- **Monolithic Humanoid Control**: Autonomous humanoid robots (such as martial arts demonstration humanoids developed in China) operate as a single unified system. When a humanoid extends one arm during dynamic kung fu maneuvers, the central controller automatically adjusts the opposite arm to offset rotational momentum and stabilize the center of mass within a closed internal loop. - **Bio-Robotic Asymmetry**: A prosthetic limb must continuously interpret and adapt to a completely separate, sovereign biological entity whose nervous system operates independently. - **Federated Sensor Grids**: The limb incorporates electromyography (EMG) arrays to read localized muscle activation prior to mechanical movement, paired with physical pressure matrices distributed across the socket and foot. - **Out-of-the-Box Stability to Continuous Adaptation**: The prosthesis relies on its foundational locomotion model to deliver safe, highly capable balance immediately upon unboxing, then uses federated edge sensing to continuously learn, anticipate, and personalize its control policies to the specific user's unique gait patterns over time.

[16:03](https://www.youtube.com/watch?v=49VC0ofQaO4&t=963s) The High Stakes of Locomotion and Prioritizing Lower Limbs

In humanoid robotics demonstrations, researchers often show robots being shoved or kicked in the snow to prove balance recovery, but human locomotion carries far higher life-or-death consequences. For a 68-year-old diabetic amputee, a single slip and fall in the snow can lead to a fractured hip, which is often fatal: "That's the end of you." Because the human cost of failure is catastrophic, assistive physical AI demands an entirely different order of magnitude in accuracy, reliability, and real-time environmental comprehension. Orchestrating these movements requires mapping and reacting to a vast plethora of terrain and biomechanical patterns.

Addressing why the focus is on lower extremities rather than hands, the division of biomechanics and market demand makes legs the immediate priority. Upper-limb amputations represent only about a quarter of all amputation cases worldwide. While hands do not bear body weight, they involve a high density of micro-actuators and complex dexterity. To create immediate, scalable impact, Bility is focusing where the clinical need is largest and mechanically feasible today, with plans to expand into upper limbs later as the technology matures.

- **Human Risk Profile**: Unlike robust experimental robots, an elderly or diabetic amputee falling on slick surfaces faces devastating complications, such as hip fractures, requiring near-zero-fault reliability. - **Prosthetic Hand vs. Leg Differences**: - Upper-limb amputations account for roughly 25% of the total amputee population. - Hands require numerous intricate actuators without bearing heavy loads. - Legs bear total body weight and are essential for basic mobility, making lower-limb physical convergence the urgent market entry point.

---

[17:06](https://www.youtube.com/watch?v=49VC0ofQaO4&t=1026s) An Open Ecosystem: The Locomotion OS and Simulator

Restoring natural mobility and dignity to amputees is a challenge far larger than any single enterprise can solve alone. Drawing an analogy to the proverb that it takes an entire tribe to raise a single child, Yishay Falick emphasizes that helping amputees worldwide regain their physical autonomy requires the collective effort of the broader technology and healthcare ecosystem: "Maybe the whole of us, maybe the whole of the world."

To drive this shift across the industry, Bility is establishing itself as a platform enabler rather than an isolated hardware developer. The company’s business and operational model centers on commercializing the foundational locomotion dataset, the real-time locomotion operating system, and the bio-robotic simulation engine. By partnering with external R&D centers, supporting internal innovation, and licensing technology to third-party manufacturers, Bility aims to foster an expansive ecosystem of next-generation adaptive mobility applications.

```mermaid flowchart TD accTitle: Bility Foundation and Partner Ecosystem accDescr: Diagram showing Bility's core technology pillars enabling internal and external R&D partners to deliver real-world applications.

subgraph Platform["Bility Physical AI Foundation"] OS["Locomotion Operating System"] Sim["Bio-Robotic Simulator"] Data["Locomotion Dataset"] end

subgraph Innovation["Development & Distribution"] Core["Internal Core R&D"] Ext["External Partner R&D"] Third["Third-Party Licensing"] end

subgraph Impact["Clinical & Consumer Market"] Apps["Adaptive Prosthetics & Mobility Applications"] end

Platform --> Innovation Innovation --> Impact ```

- **Core Commercial Offerings**: Licensing and distribution of proprietary motion datasets, the locomotion operating system, and the motion simulator. - **Ecosystem Architecture**: A hybrid R&D strategy combining internal core innovation hubs with third-party collaborations and external research integration. - **Mission**: Serving as the foundational platform layer that enables medical and robotics manufacturers worldwide to restore human dignity, fluidity, and athletic grace.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=49VC0ofQaO4&t=4s) data first, model second, hardware third — Yishay Falick

> [01:35](https://www.youtube.com/watch?v=49VC0ofQaO4&t=95s) no cloud in that loop, no human in that loop, no second take. — Yishay Falick

> [05:03](https://www.youtube.com/watch?v=49VC0ofQaO4&t=303s) your gait is not my gait and your gait tonight is not your gait this morning. — Yishay Falick

> [06:42](https://www.youtube.com/watch?v=49VC0ofQaO4&t=402s) Real data teaches the model what is true. Simulation teaches it what is possible. — Yishay Falick

> [09:07](https://www.youtube.com/watch?v=49VC0ofQaO4&t=547s) a living person with opinions. — Yishay Falick

> [10:53](https://www.youtube.com/watch?v=49VC0ofQaO4&t=653s) Millions of people will never need a robot that walks for them. They need one that walks with them. That is the harder problem. You do not own the whole body. You do not get a second take. And being wrong is not a blooper reel. It is a person on the floor. — Yishay Falick

> [16:03](https://www.youtube.com/watch?v=49VC0ofQaO4&t=963s) That's the end of you. — Yishay Falick

> [17:06](https://www.youtube.com/watch?v=49VC0ofQaO4&t=1026s) Maybe the whole of us, maybe the whole of the world. — Yishay Falick

## Mentioned Yishay Falick, Bility, Ai4, NVIDIA, NVIDIA Inception, ImageNet, ROS

## Video description Yishay Falick - Co-Founder & CEO, Bility

The Architecture of Motion: Imagine a world where an artificial limb doesn't just replace a part, but restores human dignity, fluid intention, and athletic grace. This is the frontier of physical AI, bridging advance Bio-Robotics with human biology to solve a massive global mobility crisis for veterans and diabetic amputees alike. Attendees will discover how a data-driven locomotion operating system and simulator translates complex intent into real-time adaptive motion. Join us to see how we are backing the next Era of physical convergence and how you can join this motion revolution.

[Source](https://www.youtube.com/watch?v=49VC0ofQaO4)

中文简报

Bility利用仿生机器人运动AI降低截肢者跌倒风险辅助仿生机器人需要专门的物理AI基础以在数十毫秒内预判人类意图，因为与具备全身控制能力的自主类人机器人不同，假肢平衡失误会导致灾难性伤害，而非仅仅是软件失误花絮。 Yishay Falick对具身智能领域普遍存在的“完备性谬误”提出质疑，将自成一体的自主类人机器人与辅助AI在同人类共享肉身时所需的非对称协作关系进行了对比。在低于100毫秒的严苛生理响应极限下，仿生假肢必须预判意图、适应代偿步态并防止灾难性跌倒。为解决缺乏标准化运动数据集的难题，Bility将来自多样化患者群体的真实经验采集与NVIDIA加速物理模拟相结合，生成了PB级的极端案例训练数据。Falick展示了Bility的开放平台愿景，该平台由人类运动基础模型、身体感知操作系统以及先进的超材料硬件组成，旨在成为所有协同具身AI应用的共享基础设施。与人形机器人跌倒仅被记为软件 bug 或花絮不同，助残实体 AI 面临骨折等不可逆伤害的高风险；Bility 以“数据优先、模型次之、硬件第三”为核心，通过动捕实验室与物理约束模拟器应对挑战。 (0:04) 不同于掌控全部执行器的自主机器人，仿生假肢必须解决不对称的“人机协作难题”，边缘 AI 必须在人体神经系统约 100 毫秒感知延迟前，于数十毫秒内完成感知、推理与驱动。 (1:35) 全球行动力危机主要源于慢性病，其中糖尿病致截肢占比达 50%，下肢截肢占 75%；传统静态、无传感器的刚性接受腔无法随人体组织变化自适应，形成了亟待弥合的“尊严鸿沟”。 (3:24) 针对人类步态长期缺乏通用基准数据集的瓶颈，Bility 结合真实补偿步态与基于 NVIDIA Inception 的物理仿真生成 PB 级数据，打造人体运动基础模型与身体感知操作系统。 (6:42)

Reading summary

Yishay Falick of Bility argues that human-assistive physical AI faces an asymmetric "partnership problem" that is fundamentally more demanding than autonomous robotics, requiring sub-stride coordination with a biological body where failure leads to devastating physical injury rather than a software bug. The Asymmetric Partnership and Latency Deadlines (1:35) Autonomous humanoid robots operate under what Falick calls the "completeness fallacy": a single neural policy owns every mechanical actuator, calculates a unified center of mass, and coordinates intent internally. When a laboratory humanoid falls, the error is logged as a software bug. In contrast, an assistive prosthetic controls only a single limb while remaining blind to the wearer's biological spine, hip, and shoulder dynamics. Locomotion becomes an asymmetric partnership where the artificial limb must continuously predict human intent and stabilize gait in real time. This interaction is bound by strict physiological deadlines. While an average human walking stride lasts roughly 1.0 second, biological feedback requires approximately 100 milliseconds just to register an unexpected stumble. Onboard edge AI must sense terrain, infer biological intent, and actuate motor torque within tens of milliseconds across an estimated 10,000 daily steps. The system must operate without cloud access or human intervention, ensuring absolute reliability over years of use in rain, loose gravel, and irregular terrain. Limb Loss Realities and the Dignity Gap (3:24) According to source figures, approximately 2.25 million Americans live with limb loss, a population projected to reach 3.5 million by 2050. Lower-limb amputations account for three out of four cases. While public perception often links amputation primarily to military combat trauma—which represents one in eight cases—diabetes is the primary driver, accounting for 50% of all amputations. Falick contends that conventional prosthetics impose a fundamental "dignity gap" through static, sensor-blind hardware. Standard clinical treatment fits a rigid composite socket onto living tissue that naturally contracts and swells throughout the day, creating friction, soft tissue breakdown, and high device abandonment. Furthermore, passive mechanical legs offer only a single stiffness configuration for drastically varied terrains—ranging from office floors to stairs and sand—while failing to monitor residual limb tissue health or prevent catastrophic falls. Solving Locomotion Data Scarcity Through Simulation (6:42) Major AI breakthroughs have historically relied on comprehensive foundational corpora, such as ImageNet for vision or structural crystallography for protein folding. Human locomotion, however, has lacked a standardized repository documenting how limbs articulate across diverse body types, surface transitions, fatigue states, and compensated pathologies. Models trained solely on idealized textbook biomechanics inevitably fail when deployed on individuals with altered, guarded movement patterns. To build this foundation, Bility prioritizes a "data first, model second, hardware third" methodology. The company gathers empirical kinematic baselines from healthy individuals, military veterans, and diabetic amputees inside a dedicated mobility laboratory, focusing on "distribution hot spots" such as misjudged curbs and micro-stumbles. Because vulnerable users cannot undergo thousands of physical laboratory falls, Bility feeds this empirical data into a physics-constrained simulation engine developed with NVIDIA Inception. The simulator enforces strict physical laws—mass distribution, rotational inertia, ground reaction forces, and metabolic energy expenditure—to generate petabytes of synthetic edge cases safely. Body-Aware Operating Systems and Open Infrastructure (9:07) Deploying physical AI onto living bodies requires dedicated system software. Bility is developing a bio-robotic operating system that treats body awareness as a primary computational primitive, operating like the Robot Operating System (ROS) but designed to accommodate an autonomous human user. This computational layer integrates federated edge sensors, including electromyography (EMG) arrays reading muscle activation prior to movement and distributed pressure grids across the socket and foot, allowing the device to adjust joint stiffness and personalize gait over time. Hardware engineering supports this framework through dynamic sockets, metamaterials that reduce weight while preserving fatigue strength, and energy-harvesting soles that recharge reserves from walking pressure. Bility prioritizes lower limbs over upper extremities because legs bear total body mass, represent 75% of amputations, and carry the most acute risk of fatal complications from falls. Rather than operating exclusively as an isolated hardware manufacturer, Bility structures its commercial model around licensing its foundational dataset, operating system, and simulation platform to third-party developers across exoskeletons, rehabilitation robotics, and assistive wearables. Bility frames assistive physical AI as shared infrastructure for co-embodied robotics, asserting that millions of individuals need machines that walk with them rather than for them. However, realizing this platform vision depends on successfully licensing the software stack across third-party manufacturers and validating that synthetic simulation models reliably prevent falls across unpredictable real-world environments.

精炼阅读摘要

Bility创始人Yishay Falick指出，辅助性物理AI必须打破自主机器人掌控全局躯体的“完整性谬误”，在毫秒级生理时限内与人类生物机体达成非对称协同；为此，Bility提出“数据第一、模型第二、硬件第三”的路径，通过步态基础数据集、物理约束仿真与身体感知操作系统，解决下肢假肢的尊严鸿沟并赋能广泛的具身人机共存场景。完整性谬误与非对称协同控制挑战 (1:35) 主流具身智能模型依赖将感知直接映射到电机扭矩的端到端架构，其成立的前提是系统能够控制每一个作动器并精确计算整体质心。Falick强调，将AI部署于截肢者时这一假设完全失效：算法仅掌控单侧人工肢体，对患者的脊柱、髋关节和肩部完全盲目，控制任务因而转变为非对称的“伙伴协同问题”。机器必须先验推断人类意图并实施预期性控制，在人体生物神经反馈察觉异样前完成误差代偿与姿态恢复。这种人机协同受到严苛生理时限的制约。人类典型步态周期约为1.0秒，神经系统感知踏空等机械异常需要约100毫秒，板载边缘AI必须在数十毫秒内完成感知、推理与扭矩输出的全流程。与实验室双足机器人摔倒仅作为软件漏洞记入日志不同，老年截肢者摔倒往往导致髋部骨折乃至永久丧失自理能力。该系统必须在无云端依赖和人工接管的情况下，保障每日上万步、连续二十年面对复杂地面的近乎零故障运行。截肢现状、尊严鸿沟与下肢优先策略 (3:24) 据演讲提供的数据，目前全美约有225万人承受肢体缺失，到2050年预计将达350万，其中下肢截肢占四分之三。公众常将截肢归因于战场创伤，但战斗和物理创伤仅占八分之一，高达50%的病例由糖尿病等慢性病引发。现行护理标准存在显著的“尊严鸿沟”：刚性接受腔无法适应生物组织在白天的肿胀与收缩，引发慢性疼痛和软组织损伤；单一机械刚度无法适配办公室、楼梯和沙滩等多变环境；传统设备缺乏病理组织监测能力且外观冰冷，导致极高的设备弃用率。在研发路径上，Bility明确将下肢置于手部之前。Falick解释称，全球截肢病例中上肢仅占约25%，假肢手虽然需要密集精细的微型作动器和高难度灵巧控制，但并不承载人体重量。下肢则必须承担全身负重与基础移动，跌倒后果具有毁灭性，因此解决下肢运动不仅是临床需求最大、最为迫切的痛点，也是工程上优先实现规模化突破的切入点。突破步态数据瓶颈与物理约束仿真 (6:42) Falick指出，正如ImageNet之于计算机视觉、开放互联网之于自然语言处理，AI领域的重大飞跃均依托于标准化的基础数据集，而人类运动学此前从未建立涵盖多样体型、复杂地形、疲劳状态及代偿病理的统一标注数据集。为此，Bility确立了“数据第一、模型第二、硬件第三”的体系，在专业移动实验室中采集健康人基准步态，并重点采集退伍军人、糖尿病患者在应对接受腔疼痛和疲劳时的代偿步态，避免让算法陷入理论化步态的盲区。鉴于高危人群无法在实验室中承受成千上万次真实跌倒测试，Bility在NVIDIA Inception技术支持下开发了受物理规律约束的运动仿真平台。该引擎严格遵循质量分布、转动惯量、地面反作用力与生物代谢能耗等力学边界，以实验室经验数据为基准，合成了数PB涵盖路阶误差、湿滑瓷砖和微绊倒等边缘场景的高危步态数据。Falick将其总结为：真实数据界定客观现实，物理仿真穷尽边界可能。身体感知系统架构与广泛具身生态 (9:07) 为落地运动模型，Bility正在开发一套将身体感知列为底层原语的生物机器人专用操作系统，其设计逻辑借鉴了移动机器人领域的ROS，但新增了对具有主观意图之活体人类的实时建模。在物理层面上，系统整合环境感知与肌电图（EMG）传感器阵列，提前读取肌肉意图与地形落差，通过个性化预测引擎调用底层基础运动模型，驱动自适应关节与形状可变假脚实现亚毫秒级作动。硬件配合超材料结构以平衡轻量化与抗疲劳度，并利用步行动能进行能量收集。 Falick指出，协同具身智能不仅限于假肢，还将泛化至工业外骨骼、康复机器人和动力装甲等与人体直接共承负载的领域，假肢是其中对延迟要求最苛刻、失败代价最高的使用场景。Bility并不定位为封闭的硬件制造商，而是通过对外开放其基础步态数据集、生物运动操作系统及物理仿真引擎，联合第三方研发中心与医疗设备厂商，共同构建适应多样化需求的下一代人机协同生态。 Falick强调辅助物理AI的终极目标是与人类“一同行走”而非“代替行走”。然而，该技术方案能否彻底弥合临床痛点，仍取决于电极对肌电信号长期推断的稳定性、异质软组织与机械界面的耐受度，以及其仿真数据在现实异构场景中的泛化表现；开放平台模式的推进亦依赖整个医疗与机器人产业链对标准化软件架构的采纳与验证。

Detailed notes — zh

## Background 本视频由 Ai4 频道于 2026 年 8 月 15 日在 YouTube 平台发布。内容记录了 Bility 联合创始人兼 CEO Yishay Falick 的主题演讲，探讨了如何融合物理具身智能（Physical AI）与先进生物机器人技术，通过数据驱动的运动操作系统和仿真平台将人类意图转化为实时自适应动作，从而解决糖尿病截肢者与退伍军人的行动危机并重塑生命尊严。 ## TL;DR 演讲深入剖析了人机共生系统与传统人形机器人的本质鸿沟，指出假肢面对的是无法完全支配的人体，必须在生物神经反应前的数十毫秒内实现零容错闭环自愈。为攻克人体运动领域的数据荒漠，Bility 坚持“数据第一”战略，通过移动实验室采集真实代偿步态并结合物理约束仿真器生成海量数据，以此训练底层运动大模型。团队进一步打造了将人体感知作为一级原语的运动操作系统（Locomotion OS），并优先攻克占据绝大多数承重需求的下肢场景。Bility 旨在通过开放数据、系统与仿真工具赋能整个行业生态，实现从“替人行走”到“与人同行”的范式转变。 ## Atmosphere & style 整场演讲呈现出硬核工程理性与深沉人文关怀的深刻交融，兼具学术严谨性与社会使命感。Falick 语言沉稳克制而又观点犀利，开篇将实验室机器人的跌倒花絮与截肢长者摔跤的生死风险进行对比，瞬间拉紧了叙事的张力与共情。从底层动力学模型、物理仿真架构到重拾行走自由的具象场景，内容层层推导、情绪自然升华，带给听众一种既冷静扎实又充满温度与力量的技术洗礼。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=49VC0ofQaO4&t=4s) Bility 正在构建一种必须与人类共用同一副躯体的 Physical AI（物理具身智能），其核心挑战在于必须在几十毫秒内做出精准决策，且容错率为零。 - [00:04](https://www.youtube.com/watch?v=49VC0ofQaO4&t=4s) 传统人形机器人与假肢存在本质鸿沟：机器人拥有全身 100% 的控制权且跌倒仅被视为可调试的 Bug，而假肢仅接管单一肢体，老人跌倒则可能导致骨折甚至危及生命。 - [01:35](https://www.youtube.com/watch?v=49VC0ofQaO4&t=95s) 人机共生系统的核心是协作伙伴问题（Partnership Problem），系统必须在人类生物神经反应（约 100 毫秒）之前完成感知、预测与姿态自愈调整，且完全处于离线状态。 - [03:24](https://www.youtube.com/watch?v=49VC0ofQaO4&t=204s) 全美截肢患者主要诱因是糖尿病（占比高达 50%），创伤导致的截肢仅占 12.5%，下肢截肢占总病例的约 75%，市场核心诉求是能够全心信赖的稳定性。 - [03:24](https://www.youtube.com/watch?v=49VC0ofQaO4&t=204s) 传统假肢面临严重的“尊严鸿沟”，刚性接受腔无法适应残肢随时间的微观形变，且缺乏生理感知与动态适应能力，导致大量昂贵假肢被遗弃。 - [05:03](https://www.youtube.com/watch?v=49VC0ofQaO4&t=303s) Bility 提出了由主动机器人交互接口、自主 AI 关节、环境感知与自形变足部组成的端到端全新运动架构，底层由人体运动基础模型驱动。 - [06:42](https://www.youtube.com/watch?v=49VC0ofQaO4&t=402s) 针对人体运动领域的“数据荒漠”，Bility 确立“数据第一”战略，通过真实移动实验室采集非标准代偿步态，并利用 NVIDIA 算力构建物理约束模拟器生成 PB 级合成数据。 - [09:07](https://www.youtube.com/watch?v=49VC0ofQaO4&t=547s) Bility 打造专用的运动操作系统（Locomotion OS），将人体感知作为一级原语（First-class Primitive），并结合超构材料与人造仿生骨骼进行全栈硬件工程验证。 - [10:47](https://www.youtube.com/watch?v=49VC0ofQaO4&t=647s) 具身智能的新范式不是制造一个“替人类行走”的机器人，而是打造一个能与人类“并肩同行”、在数十毫秒内化解踩空跌倒风险的人机协同伙伴。 - [13:56](https://www.youtube.com/watch?v=49VC0ofQaO4&t=836s) 在人机协同感知上，系统融合表面肌电图（EMG）神经电信号与压力传感网络，实现开箱即用的基础自平衡以及长期的渐进式个性化学习。 - [16:03](https://www.youtube.com/watch?v=49VC0ofQaO4&t=963s) Bility 采取下肢优先研发策略，优先攻克承重大、市场需求占 75% 且关乎生命安全的下肢领域，后续再拓展至执行器复杂度极高的手部义肢。 - [17:01](https://www.youtube.com/watch?v=49VC0ofQaO4&t=1021s) Bility 的商业定位是整个 Physical AI 运动生态的底层赋能者，通过开放输出运动数据、Locomotion OS 和仿真工具，驱动全行业协同创新与普及。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=49VC0ofQaO4&t=4s) 重新定义具身智能：机器与人类共用同一副躯体在人工智能蓬勃发展的当下，行业大多聚焦于生成文本的语言大模型或自动驾驶系统，但 Bility 联合创始人兼 CEO Yishay Falick 指出，他们正在构建一种全行业极少涉足的 Physical AI（物理具身智能）——它必须与人类直接共用一副躯体，并在每一步迈出时的短短几十毫秒内做出绝对精准的决策。 Falick 犀利地指出了传统人形机器人与人体义肢在容错率上的本质鸿沟：当实验室里的人形机器人摔倒时，工程师们会将其归类为一个 Bug，记录日志，甚至剪进搞笑集锦（blooper reel）中；但如果一位 68 岁的老人在走下路沿时摔倒，面临的则是髋部骨折、住院手术，甚至彻底丧失生活自理能力。两者虽然面对的是相同的物理定律与控制动力学问题，但背后的风险与 AI 范式完全不同——机器人拥有自己全身的完整控制权，而义肢面对的却是一个不受自身支配的人体。 Bility 定位于一家物理 AI 数据公司，并于当年 2 月加入了 NVIDIA Inception 初创加速计划。公司的核心研发架构包含两部分：一是捕获真实运动数据的移动实验室（Mobility Lab），采集健康人群、退伍军人及糖尿病截肢者的真实步态；二是物理约束模拟器，能将地面真值（Ground Truth）扩展为海量物理级数据。Bility 坚决贯彻“数据第一、模型第二、硬件第三”的战略优先级。 --- ### [01:35](https://www.youtube.com/watch?v=49VC0ofQaO4&t=95s) 具身智能被严重低估的挑战：从“完全控制”到“人机共生伙伴” 传统的机器人基座模型（Foundation Model）之所以能够实现端到端训练（感知输入，扭矩输出），依赖于一个隐性假设——“系统完备性”（Completeness）。机器人模型掌握全身每一个关节，清楚自身每一处质量分布，其意图与躯体完全合一。但当这种模型被迁移到人类身上时，情况发生了根本性变化：它仅仅接管了整副躯体中的一条肢体，对于人体的髋关节、脊柱、肩膀等部位既无法直接观测，也无法施加动力。这已经不再是单纯的机械控制问题，而演变成了一个高度复杂的“人机协作伙伴问题”（Partnership Problem）。机器必须提前预判使用者的意图，同时随时做好犯错的准备，并在生物神经系统察觉异常之前完成姿态恢复。人类行走的一个步周期大约只有 1 秒，而人类神经系统感知到步态异常并做出反应需要约 100 毫秒。因此，机器必须在短短数十毫秒内完成感知、推理、决策和执行全流程。 ```mermaid flowchart TD accTitle: 人机共生步态预测与毫秒级闭环控制流程 subgraph Sensing["多模态感知输入"] E["环境感知<br/>(地形与路面高差)"] H["人体态势感知<br/>(肌电/肌肉压力分布)"] end subgraph Core["核心推理与大模型"] P["个性化步态预测引擎"] M["人体运动基础大模型<br/>(物理约束与数据驱动)"] end subgraph Execution["毫秒级底层执行"] A["关节刚度与动力输出<br/>(<100ms 内完成恢复闭环)"] end E --> P H --> P P --> M M --> A ``` - **高可靠全天候要求**：全天 24 小时、连续 20 年、每天 10,000 步，在雨水、碎石路等恶劣环境以及凌晨 3 点无网络连接的离线状态下稳定运行。 - **严苛时延约束**：无云端算力介入，无人工介入回路（No human in the loop），不给算法第二次重试的机会。 --- ### [03:24](https://www.youtube.com/watch?v=49VC0ofQaO4&t=204s) 假肢领域的严峻现实：从被忽视的糖尿病患者到“尊严鸿沟” 目前全美有 225 万人承受着截肢之痛，预计到 2050 年这一数字将攀升至 350 万，其中四分之三为下肢截肢。公众普遍认为截肢主要源于战场创伤并联想到军人，退伍军人固然是核心群体，但现实中最主要的截肢诱因其实是糖尿病，占到了总截肢案例的 50%，而创伤导致的截肢仅占八分之一（12.5%）。无论是因爆炸失去下肢的 24 岁年轻士兵，还是因伤口无法愈合而截肢的 68 岁老人，他们对假肢的核心诉求完全一致——给予他们能够全心信赖的稳定性。然而，现存的假肢技术面临着巨大的“尊严鸿沟”（Dignity Gap）： - **刚性接受腔的物理折磨**：传统假肢采用单次定型的刚性接受腔，而人体残肢每小时都会发生肿胀和收缩。这种微观形变的不匹配是导致患者疼痛和最终放弃佩戴假肢的最主要根源。 - **缺乏自适应与感知能力**：设备每天紧绑在受病变侵蚀的脆弱组织上长达 16 小时，却无法监测任何生理数据，无法随办公、楼梯、沙滩或舞池等不同场景主动调节。 - **沉重、昂贵且缺乏归属感**：外观冰冷机械，让使用者感觉这并非属于自己身体的一部分，最终导致大量高价假肢被遗弃在衣柜深处。 | 维度 | 人形机器人（Humanoid Robot） | 人机共生假肢系统（Shared Bionic System） | |:--- |:--- |:--- | | **系统控制权** | 拥有 100% 关节与质量分布的完全所有权 | 仅接管单一肢体，其余躯体不可控且不可见 | | **感知与决策** | 意图与机体动作完全合一 | 必须实时推演、适应并服从人类宿主的意图 | | **容错代价** | 跌倒判定为 Bug，记录日志或进入搞笑集锦 | 跌倒可能导致严重骨折、住院与自理能力丧失 | | **实时性要求** | 允许闭环重试 | 必须在生物神经反应时延（约 100ms）前完成自愈调整 | --- ### [05:03](https://www.youtube.com/watch?v=49VC0ofQaO4&t=303s) Bility 的运动架构与端到端控制模型为了彻底打破这一困局，Bility 提出了正在申请专利的全新假肢架构。该架构不再采用冰冷的刚性套筒，而是由主动机器人交互接口、自主 AI 关节、前瞻环境感知系统、病患组织医疗监测功能、自形变足部以及可编程外观层共同构成。这种与活体组织动态贴合的接口，无法通过传统人工调参实现，必须完全由数据驱动训练。整个技术架构的运行逻辑在于：环境感知提前侦测前方地形变化，人体态势感知实时捕捉使用者的运动趋势，两者共同输入个性化步态预测引擎。因为每个人的步态都不相同，且同一个体在清晨与夜晚疲劳时的步态也存在动态差异。这些预测结果驱动底层的“人体运动基础模型”（Human Locomotion Model），在数十毫秒的窗口内输出精准的动力扭矩。在整套架构图中，感知、预测和执行模块的连线在工程上逻辑清晰，但最底层的基石——**人体运动基础模型**——却是行业内最难攻克的核心壁垒。 --- ### [06:42](https://www.youtube.com/watch?v=49VC0ofQaO4&t=402s) 攻克数据荒漠：真实实验室采集与物理约束模拟系统 Falick 回顾了 AI 发展史中的关键突破：计算机视觉拥有 ImageNet，自然语言处理拥有整个互联网的文本，蛋白质折叠预测拥有结构生物学 50 年积累的晶体学数据；然而在人体运动领域，数据却是一片荒漠，既没有统一标记的跨地形、跨身体素质运动轨迹，也没有区分健康步态与代偿性步态（Compensated Gait）的高质量数据集。没有底层数据，就根本无法训练基础模型。因此，Bility 确立了“数据即公司本身”的研发策略： 1. **真实移动实验室（Ground Truth 采集）**：采集健康人群的自然运动轨迹，同时重点采集截肢患者、退伍军人、糖尿病及血管炎患者在代偿、保护残肢疼痛部位、长途行走疲劳状态下的非标准步态数据。系统不是为了模仿教科书式的完美步态，而是为了适配具体使用者的真实状况。 2. **聚焦长尾边缘场景（Hot Spots）**：重点标注容易引发摔跤的极限分布数据，例如踩空的台阶边缘、湿滑瓷砖、身体险些跌倒却在无意识中自主平衡的瞬间。 3. **基于 NVIDIA Inception 算力的物理约束模拟器**：鉴于无法让 70 岁的老人反复摔倒 10,000 次来采集危险样本，Bility 利用真实采集的运动数据作为种子，通过高保真物理仿真器生成 PB 级（Petabytes）合成步态数据。模拟器严格遵循真实物理定律（质量分布、惯性矩、地面反作用力、生物能耗），凡是生物躯体在物理上无法做出的动作，绝不允许进入训练集。 > “真实数据教会模型什么是真实的；仿真模拟教会模型什么是可能的。”（Real data teaches the model what is true. Simulation teaches it what is possible.） --- ### [09:07](https://www.youtube.com/watch?v=49VC0ofQaO4&t=547s) 运动操作系统与全栈硬件工程生物机器人系统（Biorobotics）并非单一部件的拼凑，而是膝关节、踝关节、交互接口、传感器阵列、供电系统与有血有肉的人类大脑之间每时每刻的连续协作。为此，Bility 正在打造一套专用的运动操作系统（Locomotion OS），将人体感知作为系统的一级原语（First-class Primitive），并内置实时数据网格与物理仿真。这相当于人体外骨骼与假肢领域的 ROS（Robot Operating System），但多了一个关键维度——计算图中的核心节点是一个拥有自主意识与情绪反应的活生生的人。在硬件工程与材料层面，Bility 同样展开了深度创新： - **功能材料与超构材料（Metamaterials）**：在保持高强度的前提下实现极致轻量化，提供更强的运动稳定性，并具备可编程的外观可见性，让材料主动承担机械与生理任务。 - **一体化原型工程（Prototype Engineering）**：研发集成人造骨骼、柔性仿生皮肤、高密度传感器网格的下肢系统，并配备可通过行走压力自发电的能量收集模块。Bility 亲自制造下肢义肢硬件并非为了与传统制造巨头竞争产能，而是坚信“如果你从未亲手让一套肢体跑起来，你就没有资格声称自己拥有人体运动平台”。 --- ### [10:47](https://www.youtube.com/watch?v=49VC0ofQaO4&t=647s) 具身智能的新范式：不是替人行走，而是与人同行 Falick 强调，这套运动数据底座与操作系统的价值远超医疗器械范畴，它是所有与人类身体共享物理负荷的机器（外骨骼、康复机器人、可穿戴设备、动力装甲等）的通用基础设施。假肢只是其中技术难度最高、犯错代价最惨痛的极限应用场景。Bility 选择以开放的心态构建这一生态，因为如果行业最大瓶颈在于数据，封闭垄断只会阻碍人类福祉的演进。当前整个科技界正斥巨资教机器人如何独自在物理世界中穿行，这种探索极具价值；但世界上有数百万人永远不需要一个替他们行走的机器人，他们需要的是一个能够与他们并肩同行的机器人伙伴。 > “数以百万计的人们永远不需要一个替他们行走的机器人，他们需要的是一个能与他们一同行走的机器人。”（Millions of people will never need a robot that walks for them. They need one that walks with them.） Falick 描绘了这样一个实际场景：一位女性下台阶时未能准确看清高度，假肢系统预判落差为 6 英寸，实际落差却是 9 英寸。在接下来的短短几十毫秒内，系统自适应调整关节刚度，实现精准的能量回馈，稳稳将身体重心锁定在足底支撑面上。使用者甚至完全没有意识到危险的发生，便自然流畅地继续前行。这不仅是肢体功能的补全，更是将意图、优雅、自如以及不必时刻提心吊胆看路的自由交还给使用者，重新唤醒生命的尊严。 --- ### [13:56](https://www.youtube.com/watch?v=49VC0ofQaO4&t=836s) 问答深入：人机联合感知的神经与力学融合机制在演讲后的问答交流中，Falick 进一步通过具象案例拆解了人机协同的技术细节。他对比了国内展示的能够表演太极和功夫的人形机器人：那些机器人之所以动作极其协调，是因为感知、算力、质量分布和所有执行关节同属一个系统，移动一只手时另一只手会自然调整以保持质心平衡。但在人机共生场景下，目前尚无可以直接预测人类残余肢体下一步意图的基座模型。系统必须引入更深层次的人体态势感知能力： - **多模态物理与电生理传感集成**：利用表面肌电图（EMG, Electromyography）实时捕获肌肉运动神经电信号，预判肌肉收缩趋势，配合物理压力传感器阵列，感知压力在界面上的动态转移。 - **联邦化协同传感网络**：义肢内部的所有传感器实现去中心化的联合协作，各节点在实时广播的同时监听其他模块的需求，协同解析宿主的运动状态。 - **开箱即用与渐进式个性化学习**：截肢患者在第一次打开包装佩戴义肢时，系统就必须具备极高基线水平的自稳定控制；随后，算法会在日常行走中持续学习该个体的独特动作习惯，最终实现完全贴合宿主神经意图的自适应人机共融。 ### [16:03](https://www.youtube.com/watch?v=49VC0ofQaO4&t=963s) 严苛的可靠性要求与下肢优先策略普通机器人即使在雪地中被推倒也无大碍，但对于真实的人类患者而言，容错率几乎为零。如果一位 68 岁患有糖尿病的截肢长者在雪地中滑倒并导致髋部骨折，那可能直接危及生命。正因如此，物理 AI 系统需要达到完全不同维度的精准度、可靠性与情境理解能力。Bility 围绕这一核心诉求构建了一整套复杂的模式编排体系，确保义肢在多变且不可预测的环境中依然能够提供绝对稳定的步态支持。 ```mermaid flowchart TD accTitle: 下肢与上肢研发考量对比 subgraph LowerLimb["下肢义肢 (当前核心焦点)"] A["承重大 / 关乎跌倒与生命安全"] B["占截肢病例约 75%"] C["具备高可行性与迫切市场需求"] end subgraph UpperLimb["上肢与手部 (未来规划)"] D["非负重肢体"] E["占截肢病例约 25%"] F["执行器众多 / 机械极度复杂"] end ``` 在产品切入点的选择上，团队将注意力首先集中在下肢而非手部。手部截肢仅占全部截肢病例的大约四分之一，虽然不需要承受全身重量，但其包含极多细微的执行器（actuators），研发复杂度极高。在公司起步阶段，最务实的策略是顺应市场最大规模的痛点，优先攻克承重与行走这一具备高度可行性的下肢领域，未来再将积累的技术逐步拓展至上肢。 - **极高的容错门槛**：下肢义肢的失效不仅是机械故障，更直接关系到老年或糖尿病截肢者的生命健康。 - **截肢市场分布**：手部截肢占总病例约 25%，下肢则占据绝大多数市场需求。 --- ### [17:01](https://www.youtube.com/watch?v=49VC0ofQaO4&t=1021s) 赋能整个行业：从底层操作系统到生态协作重塑截肢者的尊严绝非单一公司能够独立完成的使命。正如人们常说“抚养一个孩子需要倾全村之力”，让全球截肢者重新找回行动自由与尊严，同样需要整个行业甚至全社会的合力。Bility 并不打算将自己局限在封闭的单一硬件制造上，而是致力于成为物理 AI 运动生态的基础赋能者（foundation enabler）。 ```mermaid flowchart LR accTitle: Bility 开放赋能生态架构 subgraph Core["Bility 核心输出"] OS["运动操作系统 (Locomotion OS)"] Sim["物理仿真平台 (Simulator)"] Data["高精度运动数据集 (Data)"] end subgraph Ecosystem["赋能外部生态"] RND["内外部 R&D 研发中心"] ThirdParty["第三方硬件制造商与开发者"] Market["多样化终端应用落地"] end OS --> Ecosystem Sim --> Ecosystem Data --> Ecosystem ``` Bility 计划通过开放与对外输出核心资产来驱动行业变革，包括对外提供高价值运动数据、运动操作系统（Locomotion OS）以及仿真器（Simulator）。通过打通核心研发中心、外部协作研发机构以及第三方商业伙伴，Bility 希望让每一位行业参与者都能从中受益，将这一底层运动智能快速带入更广泛的市场与多元化的应用场景之中。 - **商业定位**：定位为底层赋能者，对外输出数据、操作系统与仿真工具。 - **合作生态**：通过内外部研发中心联动及第三方授权，加速下一代人机感知运动技术的规模化普及。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=49VC0ofQaO4&t=4s) 当实验室里的人形机器人摔倒时，工程师们会将其归类为一个 Bug，记录日志，甚至剪进搞笑集锦中；但如果一位 68 岁的老人在走下路沿时摔倒，面临的则是髋部骨折、住院手术，甚至彻底丧失生活自理能力。 — Yishay Falick > [00:04](https://www.youtube.com/watch?v=49VC0ofQaO4&t=4s) 数据第一、模型第二、硬件第三 — Yishay Falick > [01:35](https://www.youtube.com/watch?v=49VC0ofQaO4&t=95s) 无云端算力介入，无人工介入回路（No human in the loop），不给算法第二次重试的机会。 — Yishay Falick > [06:42](https://www.youtube.com/watch?v=49VC0ofQaO4&t=402s) 系统不是为了模仿教科书式的完美步态，而是为了适配具体使用者的真实状况。 — Yishay Falick > [06:42](https://www.youtube.com/watch?v=49VC0ofQaO4&t=402s) 真实数据教会模型什么是真实的；仿真模拟教会模型什么是可能的。（Real data teaches the model what is true. Simulation teaches it what is possible.） — Yishay Falick > [09:07](https://www.youtube.com/watch?v=49VC0ofQaO4&t=547s) 如果你从未亲手让一套肢体跑起来，你就没有资格声称自己拥有人体运动平台。 — Yishay Falick > [10:47](https://www.youtube.com/watch?v=49VC0ofQaO4&t=647s) 数以百万计的人们永远不需要一个替他们行走的机器人，他们需要的是一个能与他们一同行走的机器人。（Millions of people will never need a robot that walks for them. They need one that walks with them.） — Yishay Falick > [10:47](https://www.youtube.com/watch?v=49VC0ofQaO4&t=647s) 这不仅是肢体功能的补全，更是将意图、优雅、自如以及不必时刻提心吊胆看路的自由交还给使用者，重新唤醒生命的尊严。 — Yishay Falick > [16:03](https://www.youtube.com/watch?v=49VC0ofQaO4&t=963s) 如果一位 68 岁患有糖尿病的截肢长者在雪地中滑倒并导致髋部骨折，那可能直接危及生命。正因如此，物理 AI 系统需要达到完全不同维度的精准度、可靠性与情境理解能力。 — Yishay Falick > [17:01](https://www.youtube.com/watch?v=49VC0ofQaO4&t=1021s) 抚养一个孩子需要倾全村之力，让全球截肢者重新找回行动自由与尊严，同样需要整个行业甚至全社会的合力。 — Yishay Falick ## Mentioned Yishay Falick, Bility, Ai4, NVIDIA Inception, Physical AI (物理具身智能), Mobility Lab (移动实验室), Locomotion OS (运动操作系统), ROS (Robot Operating System), ImageNet, EMG (表面肌电图), 人体运动基础大模型 (Human Locomotion Model) ## Video description Yishay Falick - Co-Founder & CEO, Bility The Architecture of Motion: Imagine a world where an artificial limb doesn't just replace a part, but restores human dignity, fluid intention, and athletic grace. This is the frontier of physical AI, bridging advance Bio-Robotics with human biology to solve a massive global mobility crisis for veterans and diabetic amputees alike. Attendees will discover how a data-driven locomotion operating system and simulator translates complex intent into real-time adaptive motion. Join us to see how we are backing the next Era of physical convergence and how you can join this motion revolution. [Source](https://www.youtube.com/watch?v=49VC0ofQaO4)

Source quotes

[0:04] Yishay Falick: 当实验室里的人形机器人摔倒时，工程师们会将其归类为一个 Bug，记录日志，甚至剪进搞笑集锦中；但如果一位 68 岁的老人在走下路沿时摔倒，面临的则是髋部骨折、住院手术，甚至彻底丧失生活自理能力。 [0:04] Yishay Falick: 数据第一、模型第二、硬件第三 [1:35] Yishay Falick: 无云端算力介入，无人工介入回路（No human in the loop），不给算法第二次重试的机会。 [6:42] Yishay Falick: 系统不是为了模仿教科书式的完美步态，而是为了适配具体使用者的真实状况。 [6:42] Yishay Falick: 真实数据教会模型什么是真实的；仿真模拟教会模型什么是可能的。（Real data teaches the model what is true. Simulation teaches it what is possible.） [9:07] Yishay Falick: 如果你从未亲手让一套肢体跑起来，你就没有资格声称自己拥有人体运动平台。 [10:47] Yishay Falick: 数以百万计的人们永远不需要一个替他们行走的机器人，他们需要的是一个能与他们一同行走的机器人。（Millions of people will never need a robot that walks for them. They need one that walks with them.） [10:47] Yishay Falick: 这不仅是肢体功能的补全，更是将意图、优雅、自如以及不必时刻提心吊胆看路的自由交还给使用者，重新唤醒生命的尊严。 [16:03] Yishay Falick: 如果一位 68 岁患有糖尿病的截肢长者在雪地中滑倒并导致髋部骨折，那可能直接危及生命。正因如此，物理 AI 系统需要达到完全不同维度的精准度、可靠性与情境理解能力。 [17:01] Yishay Falick: 抚养一个孩子需要倾全村之力，让全球截肢者重新找回行动自由与尊严，同样需要整个行业甚至全社会的合力。

Source quotes

[0:04] Yishay Falick: data first, model second, hardware third [1:35] Yishay Falick: no cloud in that loop, no human in that loop, no second take. [5:03] Yishay Falick: your gait is not my gait and your gait tonight is not your gait this morning. [6:42] Yishay Falick: Real data teaches the model what is true. Simulation teaches it what is possible. [9:07] Yishay Falick: a living person with opinions. [10:53] Yishay Falick: Millions of people will never need a robot that walks for them. They need one that walks with them. That is the harder problem. You do not own the whole body. You do not get a second take. And being wrong is not a blooper reel. It is a person on the floor. [16:03] Yishay Falick: That's the end of you. [17:06] Yishay Falick: Maybe the whole of us, maybe the whole of the world.

Original transcript

[0:04] I want to tell you about a kind of an an AI that most of this industry has never had to build. [0:10] Not one that predicts the next word. Not one that drives a car. [0:15] One that has to share a body with a human being and be right in milliseconds on every step. [0:24] So I'm again Yishai from ability. What we build is motion. [0:29] When a humanoid falls over in a lab, we call it a bug, we file it. We put it in a blooper realel. [0:37] When a 68year-old man falls stepping off a curb, it's a fractured hip, a hospital admission, often [0:46] the end of his independence. Same physics, same control problem, completely different stakes, [0:53] and a completely different AI problem because the robot owns its entire body. The man does not. [1:01] So who we are? Builty builds artificial limbs. But understand us as an AI company. [1:08] We are a physical AI data company. Two things. A real mobility lab capturing how people actually [1:14] move healthy subjects, veterans, diabetic amputees, and a physics constrained simulator [1:21] that turns that ground truth into pabytes no lab could ever collect by hand. Since February, we're [1:28] we've been part of the Nvidia inception program. Data first, model second, hardware third. [1:35] That order is the whole strategy. So here's the challenge and I want you [1:41] and I want to be precise about it because I think it is the most underrated problem in embodied AI. [1:50] You all know what a robot foundation model is. Give a policy a body. [1:55] Train it end to end. Perception in, torque out. It works because of one quiet assumption, [2:03] completeness. The model owns every joint. It knows its own center of mass because it controls [2:11] every mass. Its intention and its body are the same thing. Now that model onto a human being. [2:20] Now it owns one limb of a body it does not control. A hip, a spine, a shoulder. [2:28] It will never see and never actuate. That is not a control problem. That is a partnership problem. [2:37] It has to read someone else's intent early and be ready to be wrong. [2:48] Again, anticipate. be wrong. Recover before biology notices. A stride is about 1 second. [2:54] Your nervous system needs a 100 millisecond just to feel that something went wrong. [3:01] So the machine has to sense, infer, decide, actuate, and be finished before that [3:09] tens of milliseconds every step. 10,000 step a day for 20 years in the rain on gravel [3:17] at 3 in the morning. No clown in that loop. No human in that loop. No second take. [3:24] So why build a company around this? Because of who is waiting. [3:30] Two and a quarter million Americans live with limb loss today. Three and a half million by 2050. [3:36] Three and four have lost a leg. And here's what people get wrong. Say and this room pictures [3:43] a soldier. Veterans are central to this story. We work with them. But the biggest drive driver [3:52] is not combat. It is diabetes. Half of all amputations. Only one in eight is trauma. [4:02] A 24 year old who lost a leg to a blast. A 68-year-old who lost a foot to a wound [4:08] that would not heal. They need exactly the same thing from us. To trust the leg. [4:17] Now what we ask them to trust a rigid socket shaped once on a limb that swells and shrinks [4:24] every hour that one mismatch cause more pain and more abandonment than anything else in this field. [4:31] Not the knee, the place where the machine meets the person. One configuration for an entire life. [4:39] Office, stairs, beach, dance floor. too heavy, too expensive, strapped to a diabetic limbic [4:46] 16 hours a day on the exact tissue the disease is attacking and it measures nothing [4:53] and it does not look like it belongs to them. That is not a technology gap. That is a dignity gap. [5:03] So here's what we think the answer looks like. Our conceptual model, I won't walk [5:09] every call out. Just notice the shape of it. an active robotic interface instead of a rigid cap. [5:17] Autonomous AI joints. Environmental sensing so the limb knows what it is about to step on. [5:27] Medical function so the device finally acts on the disease it is sitting on. A shape-shifting [5:32] foot programmable visibility because how it looks is why devices end up in the closets. This is our [5:40] patent pending architecture. And notice not one of these can be handtuned. An interface that grips [5:47] living tissue has to be learned from data. Which brings us to the hard part. Here is the model. [5:56] In one picture, environment sensing tells it what is coming. Human awareness tells it [6:04] what the person is doing and about to do. Those drive prediction and personalization because your [6:13] gate is not my gate and your gate tonight is not your gate this morning. Underneath sits a trained [6:22] human locomotion model. Out of it comes actuation. Inside that millisecond window. Every arrow [6:31] on this diagram is easy to draw. Exactly one box is hard to build. It is the one at the bottom. [6:42] Every field they had transformed had a moment when the data showed up. Vision had [6:48] imageet language had the internet. Protein folding had 50 years of crystalallography. [6:55] Human locomotion has none of it. No unified tagged record of how limbs actually move. Not across [7:04] bodies, not across ter. not across fatigue and not across the difference between a healthy gate [7:10] and a compensated one. You cannot build a foundation model on that which is why nobody has. [7:18] So we decided the data is the company. A real mobility lab ground truth from healthy subjects [7:29] to learn what natural motion is and from amputees, veterans, diabetic patients, vasculitis patients [7:37] to learn what a body does when it is compensating, guarding a painful socket, [7:44] tired at the end of the long day. Because a limb does not need to imitate a textbook gate. It needs [7:52] to serve this person gate. And we tag the hot spots, not the easy middle of the distribution. [7:59] The misjudged curb, the wet tile, the stumble you recovered from without ever knowing you stumbled. [8:09] rarely any data set and the entire reason anybody falls [8:16] because you cannot ask a 70-year-old to walk into 10,000 curbs, a simulator that [8:22] takes real captured motion and generates what no lab could ever physically produce [8:30] pabytes of it. And Rule makes it worth anything at all, physics. Real mass, [8:38] real inertia, real ground traction reaction, real energy cost. If a body could not do it, [8:48] it never enters the data set. That is what we are building on the Nvidia Inception stack. [8:55] Real data teaches the model what is true. Simulation teaches it what is possible. [9:07] Then you have to let other people build on it. Because a biiobotic system is not one model. [9:13] It is a knee, an uncle, an interface, sensors, a power system, [9:19] and a human being. All having to agree continuously about what is happening right now. [9:28] So we're building an operating system for that. body awareness at the first class primitive [9:35] a real-time data grid simulation built in. Think of what Ross did for robotics. Then [9:44] at the thing Ross never had to consider one node in the graph is a living person with opinions. [9:55] Two more groups quickly because this is not only software. Functional material and meta material. [10:04] Very low weight while keeping strengths. Better locomotion stability. Programmable visibility. [10:11] Materials that do a job rather than just hold a shape. [10:19] And prototype engineering. skeleton, skin, sensor grids, a power source [10:26] that recharges from the pressure of your own walking. We're building a real lower leg limb. [10:34] Not to outmanufacture the incumbents, but because you cannot claim to have a human motion platform [10:42] if you have never had to make one walk. So why care if you're not in medical devices? The whole [10:53] I think crowd here is mostly in a specific AI tools. So why care? [11:01] Because this is the general problem of machines that share physical load with human bodies. [11:08] Exoskeletons, rehabilitation, robotics, wearables, hard suits, anywhere a person's intent and a [11:16] machine's actuation have to agree in real time. A human motion corpus and a body aware operating [11:23] system are not prosthetics infrastructure. They are embodied AI infrastructure. Again, [11:33] a human motion corpus and a bodyware operating system are not prosthetic infrastructure. [11:40] They're embodied AI infrastructured. Prosthetics is simply the hardest version and the one [11:47] where being wrong hurts the most. And we open it up. If the bottleneck in this field is data, [11:56] then hoarding the data keeps the field stuck. We do not have the moral room for that. [12:06] This industry is about to spend a fortune teaching machines to move through the world alone. [12:13] The work is real and I'm glad it's happening. But millions of people will never need a robot [12:19] that walk for them. They need one that walks with them. I repeat it. Millions [12:28] of people will never need a robot that walks for them. They need one that walks with them. [12:36] That is the harder problem. You do not own the whole body. You do not get a second take. [12:43] And being wrong is not a blooper real. It is a person on the floor. [12:50] One second in life, a woman steps off a curb. She did not fully see the limb predicts 6 in. It is [12:59] nine. In the next few tens of milliseconds, it adjust stiffness, returns energy, [13:05] holds her center of mass over the foot and she keeps walking. She never knows it happened. [13:14] An artificial limb should not just replace what was lost. It should give back intention, fluidity, [13:23] grace, the freedom to stop thinking about the ground. That is what we all call dignity. [13:32] And the amputees were, you know, could be any one of us, our parents, our friends, [13:38] the veterans that keep us safe, they need that dignity. That is the architecture of motion. [13:46] Come and build it with us. I thank you and I'm open for questions. [applause] [13:56] So again I was trying to emphasize one thing [14:00] when you have a complete humanoid robot it controls everything. It knows where the center [14:07] of mass is because it is one thing the perception and the whole movement. That is why we could see [14:15] the uh show in China that they could do kung fu because they are one piece. Their perception, [14:24] their operating system, everything is them. That is why when they move a hand, they move [14:30] the other hand to balance. You see them how they balance themselves because they are one. But now [14:39] there's no foundation model that anticipate with a robotic limb what the body is going to do. [14:46] So it has to be or has to get some human awareness skills. It need to have EMG. [14:55] Yeah. Electric miography that tells you where the muscles are going to move. So it needs to have all [15:02] those physical sensors to sense where pressure is going to happen. It needs out of the box first day [15:11] when the ampute opens the box and just get it in. It needs to be very good. But then again it is [15:19] not personalized yet to that person and then it learns to anticipate that person. Everything is [15:27] orchestrated. All the sensor are federated. They're all speaking together. They're all [15:33] speaking and others are listening, picking up the elements that they need. And then they understand [15:41] the human that carries them. Then they understand what they need to do because again [15:48] in that humanoid robotics, the whole field is one. And here we're two separated things that need to [15:57] comply complete and whenever this robot falls really nothing really happens. [16:03] You see how they push them in the snow. So what? But if you are a 68 year old man that falls in [16:11] the snow and you're diabetic, you'll fracture your hip. That's the end of you. So we need [16:18] a a totally different level of accuracy of reliability of understanding. So that is the whole [16:26] plethora of patterns that we have surrounding that thing. How we orchestrate this whole [16:33] thing was it hands hands is a totally different area. It's only about a quarter of amputations. [16:43] It doesn't carry weight. it has a lot of actuators and we think at that time point when we're just [16:51] uh beginning then we need to go where the market is and to do something that is feasible right now [17:01] later on these things will happen what I didn't said I said it only roughly [17:06] this is not one company's work this is you know know that saying how many uh people are [17:13] required ired to raise one kid. You know, it's a whole tribe. So, how many people required to [17:25] get the amputees regain their uh dignity? Maybe the whole of us, [17:31] maybe the whole of the world. So, we own them and ourselves this kind of thing. [17:37] So, we want to collaborate with many companies. We're going to uh introduce the data. That's [17:44] what we're going to be selling. We're going to be selling the uh uh operating system, the simulator, [17:49] and we will be enabling our collaborators. There's going to be a core R&D centers. It's going to be [17:54] external R&D centers. Some will be sold to third parties. Everyone needs to enjoy it. We want it to [18:00] get to the market with so many other applications. So we're we'll be the enablers of this foundation