Solving for General Robotics

Robotics has come a long way since robot arms and industrial assembly. With $19B flowing into the sector in H1 of 2026 alone, that funding will be going towards two remaining frontiers: spatial intelligence and wide deployment.

Written by: George Patin | Research Contributor, London Venture Capital Network

New Possibilities, New Constraints

For fifty years, robotics was a control problem.

Machines did exactly what they were programmed to do, operating in constrained and clearly defined environments, with progress measured in reliability and repeatability. But the paradigm has changed. As with many other sectors, advancements in AI are spilling over into robotics, enabling a completely new way of approaching how robots learn, understand and interact with their environment. Now, the open question is how to get to general machines.

Generalisation turns out to be a data problem, and the data in question — paired sensorimotor experience, the record of what a robot saw, did and felt — does not exist at internet scale. There is no Common Crawl of touch. Every hour must be earned through teleoperation, simulation or deployment, and the industry's competitive structure is reorganising around who can earn it fastest. Tens of millions of hours are already being collected, and an order of magnitude more will likely be needed before we start seeing advertisements for household robots at your local subway. In this way, deployment will lead to better robots which leads to wider deployment -- on and on.

This issue maps the reorganisation: the technology arc, the sharpest repricing in the sector's history, the companies on each side of the brain–body divide, and what robotics will look like through the 2030s and beyond.

From Control Loops to Foundation Models

The technical arc runs in three regimes.

Classical robotics encoded human knowledge as explicit models — kinematics, trajectory planning, PID control — and produced machines that were precise, but not particularly versatile. The deep-learning era (2016–2022) improved on perception but kept policies narrow: one task, one robot, one environment. The third regime — vision-language-action (VLA) models — now takes a page out of LLMs: pre-train a large foundation model on highly diverse, versatile data, fine-tune it for specific scenarios, and deploy the system across a wide range of environments.

For now, there are two critical takeaways.

First, robotics scaling laws reward diversity, not just pure volume: performance scales with the number of distinct environments and objects in the training data more than with raw demonstration count. Second, the accumulated data is transferrable: Physical Intelligence reports that 97.6% of π0.5's training data comes from robots other than the one it controls.

Together, these point to the likely winning strategy: the ability to collect broad, heterogenous data across a wide variety of environments and tasks. Which is precisely why deployment will define who gets ahead.

The Funding has Caught Up

Robotics funding is a similar story to other sectors after the 2021 peak and decline. The sector dipped from $14.7bn in 2021 to just $6.9 in 2023. That has now doubled, and is expected ot come in above the 2021 peak as the year wraps up.

Where is that funding flowing?

Along a single divide, one likely to define the next five years: better cognitive architecture, and better physical platforms. In short, brains and bodies. On one side, embodiment-agnostic foundation-model labs — Physical Intelligence, Skild AI — selling intelligence that runs on anyone's hardware. On the other, vertically integrated humanoid makers — Figure, Apptronik, 1X, Germany's NEURA — betting the data flywheel only spins if you own the fleet that generates it.

The major lingering question, though, is cash flow. The entire industry is so far running on option pricing, betting that general robotics will begin to comprise a growing percentage of the world's labor force. That is not without grounds: there are nearly 5 million industrial robots currently deployed worldwide, with potential to deploy far more as countries like the US look to reindustrialise and spur local manufacturing. What space will be occupied by universal humanoid robots, though, is for now anyone's guess.

Unitree: Looking East

While Western humanoid companies raised billions against future deployments, one Chinese company has slowly but steadily built the industry's first profitable business. Unitree sold roughly 5,500 humanoids in 2025, grew revenue to ¥1.71B (~$250M) from ¥392M a year earlier, and posted its first profit (roughly $90M) while cutting its average humanoid price from ~$85,000 in 2023 to ~$25,000 in 2025. Gross margin improved to nearly 60% as prices fell.

On 3 July 2026, the CSRC approved Unitree's Shanghai STAR Market listing — a ~$619M raise. Behind it, the Chinese capital has begun to flow into robotics more broadly: 14 rounds above ¥1B in the first half of 2026, against six in all of 2025.

The read is not that Unitree has won the race; that is nowhere near settled. So far, its robots are primarily being deployed in education, research and entertainment, not necessarily any kind of industrial labor. China, however, is definitely a case to watch. The more interesting question is whether robotics has a Moore's law equivalent. If it does, then the first companies to reach widespread deployment will be the ones with the best unit economics and a growing margin of the sector.

The Broad Cut: Who is Building What?

Below is a rundown of key names to watch, and what they are working on.

The field has broadened substantially since a decade ago, with companies playing categorically different bets. Some focus on general intelligence; others, on universal physical platforms and vertical integration. A handful, like Unitree and Fugure, are gunning for both.

For now, it is clear that the space is still wide open. We are yet to see major commercial deployments from any of the new generation robotics startups. The moat is likely in some combination of general intelligence, a flexible platform, and strong partner relationships built on trust and reliability. The brain-body stack of the first commercially deployed robot will be a critical thing to watch.

Through 2028, my guess is that value accrues to whoever converts deployment into proprietary training data fastest — not to whoever ships the most impressive hardware. The durable businesses of this cycle will be the ones whose fleets generate experience that compounds into model advantage, or whose niches are defended by regulation and integration, while undifferentiated hardware margin collapses toward a cost competitive floor.

We are still early.