HUMANOID ROBOT PAVILION STAGE

Embodied AI is entering a new phase, moving from research, prototypes, and showcase demonstrations into real-world deployment. In this session, AGIBOT will share its vision for “Deployment Year One” and discuss how physical AI systems can create measurable productivity value across industrial, commercial, educational, and service scenarios.

The session will introduce AGIBOT’s full-stack approach to embodied intelligence and its “One Robotic Body with Three Intelligences” framework, integrating motion, interaction, and manipulation intelligence into one deployable system. Attendees will learn how humanoid and multi-form robots are evolving from capability validation to scalable deployment, and how embodied AI can become a new productivity layer for the physical world. This theme builds on AGIBOT’s 2026 “Deployment Year One” positioning announced at APC 2026

Industrial automation struggles outside highly structured environments, where traditional robotics often requires extensive programming, fixed workflows, and long deployment cycles.

This presentation explores how Noble Machines is approaching industrial autonomy through a general-purpose robot learning stack designed for rapid skill acquisition and real-world deployment. At NVIDIA GTC 2026, the learning stack enabled general-purpose robots to run two distinct real-world material handling tasks using a single model trained on approximately 10 hours of demonstration data. 
The session will discuss how AI-driven whole-body control, end-to-end autonomy, and hardware-AI co-design can help robots adapt more quickly across changing industrial tasks and environments — reducing deployment friction and accelerating the path toward scalable industrial autonomy.

We believe that 2026 marks the first year of the transition from digital AI to physical AI. With the emergence of large language models and Open Claw, the embodied AI robotics industry is entering a period of significant opportunity and accelerated growth. As the first U.S. company to simultaneously deliver both humanoid robots and bionic quadruped robots, FF AI-Robotics? has already taken the lead in deploying embodied AI robotic products across real-world application scenarios in the United States.

The company has built a three-in-one robotics ecosystem platform that includes EAI Devices, centralized and decentralized robotic data collection factories, the EAI Brain, and an open-source, open-platform developer ecosystem, ultimately creating a complete commercial closed loop for physical AI infrastructure.

The robotics industry is still at the eve of large-scale industrial breakout. Frequency, necessity, and economics are among the key prerequisites for mass deployment. Today, EAI education, security inspection, logistics sorting, reception and entertainment, and robotic data collection have already proven to be viable and mature commercial application scenarios. These deployments generate data, which is then used to train the robotic brain, enabling further deployment and creating a continuous commercial flywheel.
In the future, the winners in the market may not be purely software companies or purely hardware companies, but rather physical AI infrastructure companies with strong platform capabilities and ecosystem-building power.

The promise of humanoid robots in industry has captured global attention, yet most deployments remain isolated demonstrations. The fundamental barrier is not task execution capability—it is the reliable and efficient whole-body motion that industrial environments demand. Without mastering movement first, intelligence has no physical foundation to build upon.

This session challenges the prevailing assumption that humanoid robots must immediately replace existing automation systems. Instead, it presents a counterintuitive but practical thesis: Physical AI enters industry not by taking over production tasks, but by creating value in spaces where intelligence can grow without shifting system responsibility.

Drawing from real-world factory deployments generating 2000 terabytes of physical data per month, this presentation traces the evolution of industrial operational methodology across four stages—from teach-and-repeat to autonomous physical operation—and identifies precisely where humanoid robots fit today. The conclusion is clear: the first viable application cannot be action-centric. It must be perception-centric, weakly coupled to production flow, and capable of generating continuous, verifiable data. In industrial systems, this uniquely points to quality inspection and monitoring.

Attendees will walk through the closed-loop ecosystem required to evolve physical AI from inspection toward operation: simulation environments that model physical behavior at scale, training pipelines that transform data into adaptive policies, and deployment models that align roles and responsibilities across robot providers, industrial automation partners, system integrators, and end customers. This collaborative architecture ensures clear governance boundaries, progressive capability evolution without disrupting production, and a safe path to measurable operational value.

Key takeaways for attendees:

• Why "perception before manipulation, assistance before replacement" is the pragmatic sequence for industrial adoption
• How to identify entry points where AI can demonstrate value without affecting production quality or decision authority
• What a deployment-ready responsibility model looks like across the industrial automation stack
• How the simulation-training-deployment closed loop enables continuous capability evolution from inspection to autonomous operation

The session concludes with a forward-looking perspective: sustainable physical AI evolution depends less on the intelligence of individual systems, and more on where that intelligence is placed within the broader industrial architecture.

Every major shift in industrial labor has produced an institution to govern it. The medieval guilds standardized how craftsmen learned a trade and certified their competence in marks customers across cities trusted. The factory system produced industrial training schools, time-and-motion studies, and the first standardized job descriptions. The post-war manufacturing boom built the modern certification apparatus — UL, ISO, ANSI — and the professional engineering licenses that underwrite it.

Cognitive robotics does not yet have its institutions. The robots that are now arriving on factory floors and warehouse aisles are not the robots of the past three decades. They are not pre-programmed. They reason, plan, and adapt. They learn from demonstration, from teleoperation, from continuous peer-to-peer interaction at the edge. They more closely resemble new hires than machinery.

This session discusses an institutional pattern well-suited to govern this shift, and we have used it for a thousand years: apprenticeship. Structured progression. Supervised practice. Curriculum. Assessment at stages of mastery. A credential at the end that customers across markets can trust without inspecting the work themselves.

The session translates the apprenticeship pattern into concrete operational practice for cognitive robotics, organized around five elements:

• The brain. How cognitive robots are composed today, the role of open foundation models, and what each layer means for the build-versus-buy decision.
• The curriculum. Turning operational knowledge into structured training missions across narration, demonstration, teleoperation, and experimentation.
• The examination. What a multi-stage proctored evaluation offers and what it produces — a readiness score that holds up to scrutiny.
• The journeyman phase. What deployment looks like when the robot keeps learning, including the governance and trust implications.
• The credential. What operators will need to defend cognitive deployments to CFOs, insurers, regulators, and their own line operators.

Attendees will leave with a framework for evaluating cognitive robotics offerings against the apprenticeship model and a clearer sense of which institutional pieces are emerging, which are still missing, and which the industry will have to build together.

Industrial robotics is shifting from pilot projects to real production deployment — but system validation and iteration in live environments remain slow, costly, and high-risk.

Simulation provides a faster path.

By reconstructing production environments, teams can train policies, test edge cases, and validate system readiness before deploying hardware on the line.

This talk focuses on evaluation as the key bottleneck to scaling deployment. We introduce Lightwheel’s evaluation platform, RoboFinals, which runs robot policies across large-scale simulated scenarios to identify failure modes and quantify readiness prior to production rollout.

Once validated, systems are deployed on high-frequency tasks, with real-world performance feeding back into continuous improvement.

This workflow enables manufacturers to move from fixed automation to adaptive robotic systems. The approach is being applied with partners including Analog Devices and PeritasAI across industrial and healthcare environments.

We will discuss the importance for sensors and tactile perception when it comes to human and nonhuman-like robot performance.

As humanoids and other autonomous mobile robots move from pilot programs to real-world deployments, the industry faces a challenge: AI capabilities are advancing faster than safety frameworks. Traditional safety approaches were designed for predictable environments, relying on onboard sensors and self-contained ?programming. But robotics operating in dynamic, human-populated spaces demand something more: outside-in safety. 

Onboard sensors alone cannot see around corners, through walls, or across zones, leaving critical gaps that no amount of on-robot intelligence can fully close. This session explores a holistic, proactive approach to robot safety and maximizing robot productivity using NVIDIA Outside-In Safety Blueprint, running with IGX Thor.

Join FORT Robotics and NVIDIA to discuss an outside-in approach that extends safety to include visual AI agents running on building-mounted camera infrastructure.

This session explores how Industrial Physical AI is reshaping robotics beyond narrow, task-based automation toward adaptable systems built for real-world deployment. We’ll focus on two key themes: first, how Physical AI platforms are evolving beyond single-task automation to enable broader, more flexible industrial use cases; and second, the specific design principles and capabilities required for AI-powered industrial robots—particularly in contrast to humanoid robotics. The session will conclude with practical guidance on evaluating platforms and selecting the right robotic solutions for your industrial applications.

Humanoid robots are breaking out of the lab and into real work—and the race is on to scale motion and actuation. Key trends: OEMs leaning on partners for sub-assemblies, dexterous hands driving miniaturization, and wheeled-based platforms enabling faster rollouts—with uncompromising need for safety.

This session shows how to build scalable humanoid motion systems of robotics joints, wheel-base drives, and hand actuation. We’ll show where humanoid performance targets are challenging today’s motion-control assumptions—and how that’s accelerating the next wave of new product innovation.

Key takeaways:

• Understand how to evaluate requirements for various actuations within a humanoid including joints, hands, and mobility
• Make smarter architecture trade-offs across frameless motors, precision gearing, linear actuation, wheel drives, joint modules, and brake integration
• Design for scale: prototype-to-production validation, modular sub-assemblies, and partner integration to hit performance, cost, and lead-time goal

As AI becomes embedded in safety-critical, real-world systems, the path to certification is shifting from deterministic design toward managing uncertainty, learning behavior, and evolving system boundaries. This session provides a practical, assessor-informed view on how to prepare physical AI systems for certification under emerging frameworks such as ISO/IEC 22440.

Participants will gain a clear understanding of the standard’s scope, intent, and why it matters for AI-enabled industrial and machinery applications. The session breaks down what certification bodies actually look for during assessments—highlighting key steps in the conformity process, typical documentation expectations, and the most common gaps that delay approvals.

A central focus is the distinction between traditional functional safety and AI-enabled safety. While conventional systems rely on predictable failure modes and deterministic logic, AI-driven systems introduce probabilistic behavior, data dependencies, and lifecycle challenges that demand new assurance approaches.

The session also emphasizes “design for certification” as a competitive advantage. Attendees will learn how integrating functional safety principles early—alongside structured documentation, traceability, and validation strategies—can significantly accelerate time to market while reducing certification risk.

This is a no-nonsense guide for engineering leaders, safety professionals, and product developers who need to move beyond theory and build certifiable, future-ready AI systems.

The primary bottleneck in humanoid robotics is no longer just hardware; it is the speed at which we can model complex, real-world environments for training and testing. Traditional manual modeling of facility digital twins is slow and lacks the physics-level fidelity required for precise manipulation tasks. Gatlin Robotics has developed a "Real-to-Sim-to-Real" pipeline that automates this process. By utilizing real-world video input, we generate photorealistic 3D environments and physics-ready object meshes, allowing for immediate training of humanoid control policies.

In this session, we will break down the technical architecture of this pipeline. We will discuss how we leverage the NVIDIA Isaac stack (Sim, Lab, and ROS) alongside Brev and AWS to launch these simulated worlds at scale in the cloud. Attendees will learn how automated environment generation allows for massive synthetic data collection, enabling robots to practice complex manipulation in a digital twin of a customer’s facility before the hardware even arrives on site. The talk focuses on the practical application of this technology to move humanoids out of the lab and into industrial deployment.

Most operators in 2026 aren't buying a single robot — they're managing a small portfolio of vendors. A wheeled tray-runner from one company. A pick-port stand-in from another. A folder. A packer. A greeter at the front door. Each one a separate purchase order, separate integration, separate training, separate spare-parts pipeline. The ROI on any single robot is real. The operational complexity of running five is what stops second deployments.

In this session, Carlos Oliveros Forero, Head of Product at Robot.com, will introduce R-Noid as the alternative: one humanoid platform that covers five labor categories operators can't keep staffed — picker, packer, folder, restaurant assistant, and host — with a different hat for each.

XELA’s tactile system combines uSkin sensors and uAi software to help robots understand what they touch. Each uSkin sensing point measures pressure and shear forces. The data can be processed by uAi to visualize it, detect contact points, and record tactile data in real time. Our systems can be used in automation, logistics, service robotics, and research. Tactile data enables robots to handle fragile materials and perform with precision in uncertain environments.

Deploying a humanoid robot in the real world is no longer a question of hardware alone — it's a challenge that spans data collection, model training, and sim-to-real transfer. In this session, we walk through what it takes to build a humanoid robot using NVIDIA's robotics stack from setting up your physical and simulation environments, to collecting and curating high-quality teleoperation data, to training and evaluating foundation models and world models, to deploying on physical hardware.

We'll cover what the pipeline looks like today, where the hard problems still live, and how foundation models (trained on diverse manipulation data) and world models (used for planning and sim-to-real transfer) come together to make humanoid deployment tractable.

Deploying a humanoid robot in the real world is no longer a question of hardware alone — it's a challenge that spans data collection, model training, and sim-to-real transfer. In this session, we walk through what it takes to build a humanoid robot using NVIDIA's robotics stack from setting up your physical and simulation environments, to collecting and curating high-quality teleoperation data, to training and evaluating foundation models and world models, to deploying on physical hardware.

We'll cover what the pipeline looks like today, where the hard problems still live, and how foundation models — trained on diverse manipulation data — and world models — used for planning and sim-to-real transfer — come together to make humanoid deployment tractable.

DEX is transitioning from a traditional dual-arm robot into a scalable embodied AI platform by aligning with NVIDIA’s three-layer "Physical AI" strategy. This framework utilizes a "three-computer solution": DGX-class systems for heavy-duty AI training, Omniverse and Isaac Sim for high-fidelity simulation and synthetic data generation, and Jetson Thor hardware for real-time edge inference. By leveraging this stack, DEX can move beyond fixed programming to become an intelligent system capable of perceiving, reasoning, and learning through a continuous data flywheel. The core of this strategy is a development loop where real-world data from DEX's tasks is expanded in simulation to train advanced vision and reinforcement learning models. This approach reduces manual engineering and on-site tuning, allowing DEX to adapt to diverse environments—such as varying lighting or object placements—without requiring exhaustive physical trials. Ultimately, this partnership positions DEX as a commercially scalable solution that grows more capable with every deployment.

Physical AI is a market where early decisions compound. The buyers who get the foundational bets right will get years of compounding value— and others risk spending years reworking deployments that never quite worked. In this session, Jon Battles, VP of Technology Strategy at Cobot, lays out the handful of bets every serious robotics buyer has to make— on platform, on data, on use case, on where to start and where to scale— and shares how to think through each at a strategic level. He'll then walk through the latest product updates from Cobot, and how it fits into the framework. With 40+ years in robotics including 17 at Amazon, Jon brings the long view to a fast-moving moment. Conversational, with audience Q&A.

Humanoid robots are leaving the lab. End users across manufacturing, logistics, and services are actively evaluating deployments but the safety picture remains unclear. There is no dedicated functional safety standard for humanoids yet, and the existing frameworks that apply, including ISO 10218, IEC 61508, IEC 62998, and ISO 13849, were not designed with mobile, AI-driven bipedal systems in mind.

This panel brings together humanoid manufacturers, robot safety and regulatory experts, and AI specialists to give attendees a grounded, practical view of where the field actually stands. The goal is to leave participants with a clear understanding of what is possible today and what responsible deployment looks like in the interim.

Topics will cover the current regulatory landscape and how manufacturers are navigating it, real-world safety performance of systems already in the field, key hardware and software requirements for safe human-robot collaboration, and the certification and risk assessment steps that matter most before go-live. Panelists will share direct experience designing safety architectures, advising on compliance, and building AI guardrails for physical systems operating alongside people.

The session is structured for end users, integrators, and technical decision-makers who need to cut through the noise and form a realistic view of humanoid safety - not as a barrier to adoption, but as the foundation that makes adoption sustainable.

Panelists represent Reynolds & Moore, Saphira AI, Synapticon, and a humanoid manufacturer (to be announced soon).

Social Intelligence for Homes and Workplaces explores what it will take for robots to move from impressive demos to truly useful everyday partners. This talk will look at the foundational requirements for helpful robots: perception, mobility, safety, adaptability, and, most importantly, the ability to understand human intent in real-world environments. As humanoids and quadrupeds enter factories, homes, offices, hospitals, warehouses, and public spaces, their success will depend not just on what they can do, but on how naturally they can work with people. Join us for a forward-looking conversation on the future of socially intelligent robots, and how they will reshape the way we live and work.

An airplane is not a bird. While flight was inspired by nature, aviation succeeded because engineers made deliberate design choices based on physics, materials, safety, and scalability—not biology alone. The same principle applies to humanoid robots in manufacturing. Although inspired by the human form factor, industrial humanoids must be engineered for performance, reliability, cost efficiency, and integration within complex production environments.

This educational session explores the critical design decisions required to move humanoid robots from concept to commercially viable manufacturing assets. Rather than focusing on anthropomorphic mimicry, the presentation examines how engineering trade-offs determine whether a humanoid platform succeeds on the factory floor.

Attendees will gain insight into the foundational architectural decisions that shape industrial humanoids, including actuator selection, power density, thermal management, mobility strategies, payload-to-weight ratios, and durability standards. In high-duty-cycle manufacturing environments, design priorities differ significantly from research or demonstration robots. Cycle time, uptime, maintainability, and safety certifications become central to the engineering equation. A humanoid designed for industrial deployment must withstand repetitive motion, variable payloads, and real-world environmental factors such as dust, vibration, and temperature variation.

The session will also address control system architecture and the integration of AI-driven perception with deterministic industrial controls. Unlike laboratory environments, manufacturing systems demand predictable behavior, seamless interoperability with PLCs and MES systems, and compliance with established safety frameworks. Design decisions around sensing modalities, redundancy, fault tolerance, and human-robot interaction directly impact adoption readiness.

Another focal point of the discussion will be manufacturability and scalability. Designing humanoids for manufacturing is not only about deploying robots into factories—it is also about building robots at scale. Component standardization, modular joint design, supply chain resilience, and serviceability are essential to achieving economically viable production volumes. The talk will outline how thoughtful engineering choices influence total cost of ownership, service intervals, and long-term platform evolution.

Energy efficiency and operational economics will also be examined. Battery strategy, power consumption optimization, and charging infrastructure must align with shift-based operations and productivity targets. Industrial humanoids must deliver measurable ROI, not simply technical novelty. The talk will provide a framework for evaluating design decisions through the lens of commercial impact.

Finally, the session will also explore the balance between human-inspired versatility and task-specific optimization. In many manufacturing applications, success depends on determining where human likeness adds value—such as tool compatibility or workspace accessibility—and where alternative mechanical solutions may outperform strict anthropomorphism.

Designed for engineers, robotics leaders, integrators, and manufacturing executives, this discussion will offer a practical, systems-level perspective on building humanoid robots that work in the real world. By reframing the conversation from “how human-like can we make it?” to “what engineering choices make it commercially viable?”, attendees will leave with a clearer understanding of the path from inspiration to industrial implementation.