31/622 → 58/739; +2.86 percentage points of formation share.
A striking number of these companies are not building robots at all but building the raw material robots will eventually learn from: Cortex AI, Human Archive, StableBrowse, and Asimov are all effectively data infrastructure plays, racing to capture human movement and workplace footage before anyone has settled how consent, provenance, or compensation for that footage should even work. That split shows up again in simulation, where One Robot exists specifically because physical robot time is too scarce and slow to train on directly, and in software brains like Lightberry and Efference-style embodied AI layers that assume the hardware is someone else's problem. Meanwhile the formation's own 1.59x growth looks almost sluggish next to the categories feeding it: Industrial at 8.98x, Hardware at 5.49x, and Automation at 5.19x are all accelerating far faster, suggesting capital is piling into the physical substrate and tooling around robotics faster than into robotics companies themselves. Fleet-scale deployments from Servo7 and Remy AI inside warehouses are also quietly generating a second problem nobody priced in: when a robot fails near a person or a conveyor, the failure looks like a distributed-software incident that needs synchronized sensor and operator logs to debug, and eventually a certifiable safety case to keep operating. None of this was practical before large vision-language-action models like the Pi0.5 family became usable checkpoints rather than research demos, which is roughly why 2025 is the year data-capture and simulation companies became fundable in their own right rather than being an internal team inside a robotics startup.
Robot fleets turn physical failures into distributed-software incidents, but debugging also requires synchronized sensor, action, environment, and operator context.
Map deployment-stage robotics companies against tools for fleet telemetry, incident reconstruction, regression diagnosis, and reliability operations.
Commercial robots operating near people may need continuously updated evidence that software, hardware, and operating procedures remain within an approved safety envelope.
Interview deployers and certification specialists about recurring evidence assembly, change impact analysis, simulation records, and audit handoffs.
Imitation learning and teleoperation can transform human demonstrations into durable training assets without a standard way to track permission, compensation, permitted uses, or revocation.
Trace how robotics teams acquire demonstrations and whether enterprise buyers require consent lineage, contributor rights, or usage restrictions.
More simulation does not by itself show that policies survive sensor noise, environmental variation, hardware wear, or rare real-world conditions.
Compare simulator, robotics, and deployment workflows for independent benchmarks, transfer metrics, scenario coverage, and reproducible failure suites.
Autonomous systems need repeatable evaluation across contested sensing, communications loss, adversarial behavior, mission changes, and human command boundaries.
Map program test requirements against scenario generation, hardware-in-the-loop evaluation, mission rehearsal, red-teaming, and evidence delivery.
Lambda Robotics builds and deploys robots for the infrastructure powering frontier AI.
Industrial aOS between humans, agents, and robots
Autonomous CNC factories for faster hardware
We design and manufacture high-mix drone motors and robotic actuators in LA. DJI, Apple, and Ford don't settle for OTS components. Likewise, the teams that win the Western robotics market will build seamless machines from tightly integrated software and hardware. Pendulum Robotics designs motors and actuators around your robot's unique form factor and mission - then we rapidly bring them to life with proprietary software and tooling geared for high-mix machining and assembly. Our first factory in LA can create thousands of motors per month up to 200mm OD. We're doing for motion components what 3D printing has done for structural components. Hardware teams that can build and iterate at the speed of software will create elegant, reliable, ROI-positive robots and win massive markets.
Shepherd Robotics builds general purpose robots for skilled physical work across AI infrastructure and advanced manufacturing. Our robots are powered by robot foundation models and use modular, flexible hardware designed to adapt to different environments and tasks. We train our systems on a combination of real world and simulation data so they can learn complex manipulation skills and generalize across workflows. We also provide integration into customers’ existing workflows, starting with data center construction and operations, electrical manufacturing and testing, and electronics manufacturing, where traditional automation struggles with variability, dexterity, and changing environments.
Deploy's core belief is that robots should be useful. Getting a generalist robot to do real work is fundamentally the same problem across industries. We build everything between the model and the task to bring generalist robots to the businesses that actually need them.
Every serious robotics lab is bottlenecked on the same thing: demonstration data. The field has converged on egocentric video and hand pose as the substrate, and both are now cheap to collect at scale. What neither of them contains is force. A video of someone picking up an egg and a video of someone picking up a rock are nearly identical in pixel space and nearly identical in joint angles. They differ almost entirely in a quantity nobody is recording. Policies trained without it learn geometry and inherit no sense of touch. Hemlock closes that gap at the source. Our device, the Bracelet, is a wrist-worn surface EMG band that reads the electrical activity of the forearm muscles that actually drive the hand, which means it recovers contraction force directly rather than inferring it from motion. This distinction is the whole thesis: force cannot be reconstructed downstream of kinematic
twentyfour26 builds autonomous CNC machining cells for rapid prototyping, so hardware teams can iterate with software speed.
Every instance of intelligence occurring in nature is embodied. Scientific intelligence is no different. We are building general-purpose robotic experimentalists: systems that unite scientific reasoning and physical action to iterate toward discovery by observing, hypothesizing and experimenting.
Dreamscale Labs is building the fastest cloud inference platform for robotics. As complex physical AI models become difficult to fit on expensive on-robot GPUs, we let users offload inference to the cloud and serve them with real-time latency. So far, we've hit 40-50% lower latency than existing neoclouds.
No comparable outside activity is connected yet.
These share a word with “Manufacturing and Robotics” but are tracked as separate categories. The real overlap below is the actual share of Manufacturing and Robotics’s companies also found in that category — not a guess.
Cambrian Signal, “Manufacturing and Robotics,” publication yc-2026-10-02-semantic-v8, snapshot October 2, 2026.
cambrian.jafar.com/formation/manufacturing-and-robotics