The Top Physical AI Startups in YC S26
The physical AI companies in Y Combinator's Summer 2026 batch are building machines that do real-world work and the software layer that connects AI models to them. Most already have paying users or shipped product in the field. Here are the top physical AI startups (in no particular order) from the batch.
Agency Tool Company — Shipping Software to Robots in the Field
Agency Tool Company (agencytool.com) sells the plumbing for getting code onto robots that are already deployed. Its first product, ATC Deploy, sends only the bytes that changed and picks up where it left off when a connection drops, which the company says makes updates as much as 20 times faster than a docker pull or scp. It works with OS images, Docker images or a plain file tree, and an offline robot can take an update over USB. A second product, ATC Build, runs tests on the actual target hardware as a CI service, with support for the Nvidia Jetson line, Raspberry Pi and other embedded targets.
Founders Jack Morrison and Davis Foster spent eight years at Scythe Robotics putting hundreds of autonomous mowers into commercial service before selling the company to ASI in 2026, and they describe Agency Tool as the deployment platform they wanted during that stretch. Deploy is live with launch partners in agriculture, construction and logistics robotics, including Burro, which has deployed more than 750 machines worldwide, concrete-polishing company Tempo, and inventory drone company Gather.
Aktoria Robotics — A Teleoperation Network for Robot Deployments
Aktoria Robotics (aktoria.com) runs a 24-hour network of trained robot operators across India, Mexico and the United States, paired with its own low-latency teleoperation software. The pitch in its launch post is that most robots billed as autonomous still have a person in the loop, and that standing up an in-house operations team is a slow, expensive detour for a robotics company trying to get to a first deployment. Aktoria reports 40ms cross-country latency and prices the service at less than half the cost of hiring operators directly.
Jaskaran Singh Walia and Shreenabh Agrawal are ML and robotics graduate students at Carnegie Mellon. Agrawal previously worked at Zoox and has first-author papers at robotics conferences; Walia has authored more than 15 papers and researched at MIT, Cambridge and Microsoft Research.
Robocurve — Independent Benchmarks for Robots
Robocurve (robocurve.org) is trying to establish what today's robots can actually do. Frontier labs currently evaluate their own robotics models in-house, and public evidence mostly takes the form of demo videos of narrow tasks, so there is no shared reference for where the frontier sits. Robocurve builds open-source tooling and runs reproducible benchmarks on physical hardware, with tasks that range from playing Jenga and making a sandwich to building out a data center. The team has shipped v1 of its open-source framework, Inspect Robots, and completed a first pilot scoring a frontier model on a real robot.
Founder and CEO Jay Chooi holds a BA and MA from Harvard, won a Rhodes Scholarship, and was a top contributor to Inspect Evals, the UK AI Security Institute's evaluation framework. He has published at ICML, ACL, EMNLP and ACM EC, and previously spent time at the UK AI Security Institute and MATS.
Maingen — Training Environments for Industrial Operations
Maingen (maingen.ai) builds reinforcement learning environments that mirror how industrial sites are actually run, so frontier labs can train models to operate power plants, factories and data centers. The company's argument in its launch post is that this data was never written down on the internet. It lives with operators, on control desks, and inside maintenance systems and vendor portals, so the team embeds with operators and converts their decision-making into environments a lab can train against. Its first public benchmark, SolarBench, tests whether an agent can manage a solar portfolio, and the company says it is already working with frontier labs.
Phillip Yan, the CEO, previously worked at 0x, Scale AI and Coinbase, and spent time in power trading. CTO David Yang has worked at Scale AI and Amazon and was cited by Anthropic for an ICLR 2025 mechanistic interpretability paper.
HERA — Catching Drawing Errors Before Anything Gets Cut
Every custom metal part begins as a 2D drawing, and each of those drawings gets reviewed by hand against ASME and ISO standards before it reaches a machine. That review usually lands on the most senior engineer in the building and takes days. HERA (manufacturingintelligence.org) does the check in minutes, reviewing GD&T callouts, tolerance stack-ups and title block requirements against ASME Y14.5, the relevant ISO standards and a customer's own design intent. A missed tolerance callout or an ambiguous datum reference is what turns into a scrapped part or a failed inspection, and each review adds to a record of the errors a given shop tends to make.
CEO Meera Patel has worked around her family's plants and delivered projects for Pfizer and TE Connectivity, where she built a discrete-event simulation, and published machine learning research at the University of Washington. CTO Noelle So grew up in an industrial park her family runs in the Philippines and built a water treatment facility there at 19 that added 40 liters per second of capacity. HERA works with manufacturers of data center ancillary equipment, renewable energy systems, robotics, heat exchangers, power plants and pressure vessels.
Torus — Agentic Tools for Engineering Firms
Torus (usetorus.com) builds agentic engineering tools for the firms and project owners that design capital projects, a category it describes as Legora for physical engineering firms. The company says it is working with multiple Fortune 500 companies on data centers and energy installations.
CEO Marcus Lima previously founded Heimdal, a direct air carbon capture company, and is an Oxford and Cambridge graduate. CTO Rahul Thayil previously founded Orion, which worked on AI threat intelligence and response automation.
Cosmic Robotics — Heavy-Lift Manipulation for Infrastructure Builds
Cosmic Robotics (cosmicrobotics.com) builds construction equipment for the lifting, placing and fastening still done by hand. Cosmic-1, a 10,000-pound autonomous robot, picks up 90-pound solar panels and sets them onto racking with millimeter precision in heat, freezing rain, snow and dust. It has installed tens of thousands of panels on some of the largest solar projects in the US and, the company says, more than doubled the labor productivity of the crews it works alongside. Cosmic is expanding into data center construction and is working with NASA on lunar construction robotics.
CEO James Emerick was a robotics engineer at Built Robotics, where he deployed autonomous earthmoving machines in production, and a research engineer at Autodesk Research. CTO Lewis Jones was a propulsion engineer at Relativity Space with earlier stints at SpaceX, NASA JPL and NASA JSC. The founders frame the Earth work as the path to the harder version of the problem: revenue and field data from solar and data centers train the systems they intend to send off-planet.
Together the eight sit at different points in the same stack: putting robots on jobsites, getting code onto machines, keeping humans in the loop while autonomy catches up, measuring what models can do, generating the data to train them, and checking the engineering work that precedes anything being built.