Inside the Physical AI Boom: 12 Robotics Startups

The comparative data reveals the capital division across modern robotics. While full-stack hardware manufacturers like Figure AI require hundreds of millions of dollars to build assembly plants, component and software startups are raising targeted rounds to dominate specialized layers of the stack. An emerging robotics venture no longer needs to build its own motors, program its own vision models, and code its own pathfinding scripts from scratch. Founders can license drive bases from specialized platforms, install electronic skin from MoSense, train motor policies in Machenta simulations, and deploy universal brain models from Skild AI. That modularity is accelerating development schedules from multi-year research projects into rapid commercial deployments. Specialized robot funding was analyzed in our review of DeepSeek backing robotics startup Winverse.

Industrial Muscle: Full-Stack Humanoids and Logistics Swarms

Building the brain and senses is only half the battle; the machine must still perform heavy labor without tipping over or breaking parts. These three startups are building the full-stack mechanical bodies designed to replace human physical labor:

7. Destro AI (San Francisco, California): Operating in the high-pressure logistics sector, Destro AI secured $8M in seed funding to solve warehouse fleet bottlenecks. Instead of focusing on single expensive humanoids, Destro builds coordination software that controls fleets of mobile rovers and robotic sorting arms simultaneously. The system treats an entire ten-acre distribution center as a single physical computer: balancing package intake, directing autonomous pallet jacks, and adjusting sorting speeds to match truck departure schedules. Logistics fleet coordination was detailed in our report on Destro AI securing seed funding to coordinate warehouse robots.

8. Physical Intelligence / Pi (San Francisco, California): Led by veteran roboticists from UC Berkeley and Stanford, Physical Intelligence raised $400M from prominent backers to build π0 (pi-zero): a general-purpose physical policy model. The system operates across diverse hardware bodies, learning how to fold laundry, bus tables, and assemble electronics by treating motor commands as continuous tokens. The team demonstrated that a single neural model can control distinct robotic platforms, proving that physical manipulation can be generalized across tasks without requiring bespoke code for every action.

9. Figure AI (Sunnyvale, California): Figure AI represents the most ambitious full-stack humanoid builder in North America. Raising $675M from investors including Nvidia, OpenAI, and Microsoft, Figure designs both the mechanical skeleton and the electrical actuation of its Figure 02 humanoid. Standing five feet six inches tall and weighing seventy kilograms, the machine operates on factory floors: placing sheet metal components into automotive presses at BMW production plants. Powered by custom high-torque electric actuators and on-board vision reasoning, Figure 02 demonstrates that bipedal humanoids can work alongside human technicians on active assembly lines.

Specialized Application Layers: Healthcare, Inspection, and Safety

General-purpose humanoids capture public attention, but specialized robots designed for specific commercial environments are generating verified business income today. These final three startups are proving commercial utility in demanding enterprise sectors:

10. Norbert Health (New York / Paris): Norbert Health raised $14M to deploy autonomous monitoring rovers into hospital wards and nursing homes. Built with contactless medical radar and thermal vision sensors, the machine moves quietly down hospital corridors: scanning patient vital signs, measuring respiration rates, and detecting fall hazards without touching human patients. The platform relieves exhausted nursing staff of routine hourly checks, freeing human medical workers to focus on urgent clinical care. Healthcare automation platforms were examined in our review of Norbert Health securing Series A funding for nursing robotics.

11. Safeworld (San Francisco, California): As autonomous machines enter shared human workspaces, preventing physical injuries is a legal necessity. Safeworld raised $12M to build continuous auditing and safety interception software for physical automation. Safeworld installs multi-spectral sensor perimeters around collaborative robot work cells, monitoring the velocity and proximity of human workers. If a mechanical arm swings toward an approaching technician, Safeworld intercepts the motor controller at the hardware level, cutting power within five milliseconds to prevent crushing accidents. Safety auditing systems were detailed in our analysis of Safeworld raising $12M to audit generative robotics.

12. Mach9 (Pittsburgh, Pennsylvania): Mach9 secured $34M in Series A funding to turn raw physical sensor scans into engineering models. The company software processes visual and lidar telemetry collected by mobile inspection robots, converting real-world scans of bridges, electrical substations, and rail lines into millimeter-accurate computer-aided design (CAD) files in minutes. By automating spatial data analysis, Mach9 allows civil infrastructure owners to inspect physical assets without sending human inspection crews into dangerous tunnels or high-voltage facilities. Automated infrastructure scanning was explored in our report on Mach9 securing $34M for civil engineering CAD tools.

The Actuator Deficit: Where Silicon Meets Physics

While software models and perception systems have advanced rapidly, the physical robotics boom confronts an unyielding mechanical bottleneck: electric actuators. In human biology, skeletal muscles provide high torque, compliance, and explosive power while consuming calories with astonishing energy thrift.

In mechanical engineering, replicating the power density of human muscle requires heavy copper windings, high-grade rare-earth magnets, and precision planetary gearboxes known as harmonic drives. The global supply chain for precision harmonic drives is concentrated in a handful of specialized manufacturers in Japan and Germany, creating waiting periods of up to six months for custom gear assemblies. Furthermore, powering twenty-eight electric motors simultaneously across a walking humanoid drains lithium battery packs quickly: most bipedal prototypes operate for barely two to four hours before requiring a battery swap. Startups like Figure and Apptronik are pouring capital into proprietary motor design: engineering axial-flux electric motors that generate higher torque per kilogram while shedding excess thermal heat. Until actuator power density doubles and battery chemistry advances, physical robots will remain constrained by weight and operational stamina. Mechanical supply chain constraints were examined in our review of Foundry Robotics raising $100M for autonomous factory lines.

The Data Scarcity Problem: Why Text Scraping Fails

The secondary hurdle slowing physical machine intelligence is data collection. To train a language model, research labs scraped petabytes of public text from internet web pages, feeding billions of sentences into server clusters at low expense.

There is no public internet repository of physical manipulation data. You cannot scrape the sensation of threading a needle or balancing on an icy sidewalk from a website. Collecting physical training data requires operating real machines: hiring human teleoperation workers to wear virtual reality headsets and haptic gloves, manually controlling robots through thousands of hours of repetitive tasks to record joint angles, motor torques, and visual feeds. Startups like MaxInsights and Mecka AI emerged specifically to solve this data deficit: setting up physical training warehouses where human workers guide robots through thousands of daily household chores to generate structured training datasets. The cost of collecting physical action data is hundreds of times higher than downloading text files, meaning that startups with superior simulation environments and capital reserves hold an immense advantage. Robot training data acquisition was detailed in our report on MaxInsights raising $30M to convert household chores into robot data.

The Long-Term Assessment on Physical Artificial Intelligence

The rise of physical machine intelligence marks the beginning of the most profound industrial transformation since the invention of the steam engine. For three decades, the digital economy operated as an isolated realm: creating wealth inside computer networks while leaving the physical world of roads, factories, and farms largely unchanged.

The twelve startups profiled in our review prove that the boundary separating software code from physical matter has dissolved. By combining advanced spatial foundation models with sensitive electronic skin, high-torque actuators, and rigorous safety controls, these ventures are constructing the machines that will clean our hospitals, harvest our food, assemble our electronics, and maintain our critical infrastructure. The road ahead will require overcoming mechanical bottlenecks, managing supply chain constraints, and ensuring physical safety in human workspaces. The direction of travel is undeniable: the future of artificial intelligence does not belong inside a glowing screen, but in the capable, tireless machines that walk beside us in the physical world.

Read More on TechRobust:

• Human Capital raises $100M for physical AI venture studio
• MoSense raises RMB 100M to build robot electronic skin
• Machenta raises $40M to train robots in virtual worlds