Building a City on Mars

Our mission at Cosmic Robotics is to build a city on Mars. In 2024, when we started the company, it was clear to us that while rockets were quickly maturing to enable transportation of people and cargo to Mars by the mid-2030s, there were several other core technologies that we still needed to develop for a sustained human presence there. One of these challenges to building and operating a Martian city is labor.

A city needs to complete various agricultural, construction, manufacturing, and maintenance tasks to sustain itself. Researchers have cited wildly varying minimum viable population sizes to achieve self-sufficiency, but SpaceX’s estimate is around one million people to staff all critical functions. They seem to be serious about building a city of this scale and they are sizing their factories accordingly. Their plan is to launch 1,000 Starships every 26 months for 20 years straight to achieve one million inhabitants on Mars.^[1] For reference, peak commercial airplane production by Boeing was 806 planes in 2018 while SpaceX’s Starfactory is sized for up to 1,000 Starships per year.^[2][3] While we hope SpaceX succeeds in the current baseline plan, we also see that this is an extremely challenging target. Beyond the need to raise the necessary capital and manufacture enough rockets, the long time horizon to achieving self-sufficiency means even a temporary slip in support (e.g. political or economic shock on Earth) would jeopardize the entire operation half-way through. Because of this, we believe the technologies that can radically reduce the minimum viable size of a Martian city are the next bottleneck to accelerate self-sufficiency.

Cosmic Robotics was founded to increase the productivity of workers on Mars to enable self-sufficiency with ~10,000 citizens instead of a million. If such a dramatic improvement in productivity could be achieved, not only can the entire city be staffed in a single Earth–Mars launch window, but it also allows a much smaller crew to complete most of the infrastructure deployment before the bulk of the citizens arrive. On the other hand, we need to recognize this is an extremely ambitious target. SpaceX is a ~14,000 person company today and while it’s impressively vertically integrated, it still doesn’t mine its own metals, refine its own propellant, or operate farms to feed its employees. Even SpaceX can’t be considered anything remotely close to self-sufficient in the Martian sense.

When we started Cosmic, achieving the 10,000 person city target seemed nearly impossible. However, our internal development to date and rapid advances in AI research (particularly in Physical AI) in the last two years have strongly shifted our estimate for how much of the city construction and maintenance process can be automated. We’re now confident there is a development path to prepare autonomous robots for city construction on Mars by the mid-2030s, and have identified four key technologies that are missing today. These form the core layers of our technology stack.

To date, we’ve primarily focused on building Layers 0 and 1 to unblock supervised autonomy on terrestrial jobsites and begin collecting data. Cosmic-1 has now achieved 2x labor productivity compared to manual methods for terrestrial solar farm construction. In order to replicate this success in other construction verticals and further increase productivity, we are accelerating development of Layers 2 and 3. In 2025, we became the first company to win a contract from NASA to develop edge reasoning VLMs for in-space and lunar applications to build towards Layer 2. We have also begun deploying Constellation, our jobsite planning and mission control software, on which AI agents will be deployed to coordinate tasks in the field. With recent advances in reasoning models, we are now confident Layer 3 built on Constellation is a matter of engineering execution and not frontier research.

Layer 0 — Hardware300kW-class fully electric construction-scale robots

Today, construction is primarily completed on Earth with diesel powered hydraulic equipment that will be near impossible to deploy on Mars. Mars has a thin atmosphere, less than 1% of Earth’s, and composed mostly of carbon dioxide and nitrogen. Even if carbon-based fuels were produced (such as methane for rocket fuel), equipment and robots would need to store both the fuel and oxidizer. This is both a major safety concern as well as a challenge to operating with minimal maintenance for decades.

Retrofitting existing equipment with batteries and motors may sound sufficient, but today there is almost no electric construction equipment deployed at scale on Earth. The supply chain and know-how to build construction scale batteries, power electronics, and actuators that can operate with peak power draw up to 300kW (400hp) are virtually non-existent.^[4]

Not only do we believe electrifying construction equipment and adding new classes of construction robots capable of fine manipulation is critical to building on Mars, we believe this is inevitable on Earth regardless. Internal combustion engines and hydraulics are being replaced with batteries and electric actuators across aerospace and automotive industries due to longer life and lower maintenance needs. While construction equipment has lagged behind, there is increasing demand for transitioning in the next decade.^[5] Additionally, many tasks with the highest levels of labor shortage on Earth are skilled labor tasks performed by craft professionals like electricians and plumbers that currently don’t benefit from equipment or autonomy at all.^[6] The need to maintain a large skilled workforce on Mars will be prohibitive, and we see a clear demand for new construction robot solutions to address these shortages on Earth today.

Layer 1 — ControllerSafe and adaptive high-frequency robot controller

There is a gap today between traditional industrial robot controllers and leading Vision Language Action (VLA) and World Action Models (WAM) research that must be bridged to enable deployment of advanced adaptive robots.

Almost all modern industrial-scale robot manipulation systems have minimal feedback control. Robot arms in car factories simply follow preprogrammed motions and are incapable of adapting on the fly to changing environments. This is partly why robots have had minimal adoption in outdoor environments where wider tolerances, weather, and other environmental conditions are hard to predict ahead of time.

On the other hand, the latest research in VLA and WAM that demonstrate generalizability and emergent behaviors are hampered by a lack of introspectability, debuggability, and safety guarantees necessary for real world deployment. Particularly with large industrial scale robots where robot motion can lead to damage or injuries, there is a need for safety guarantees on par with autonomous vehicles. Luckily, compared to driving 80mph on highways or near pedestrians, the technical challenges of implementing such a safety feature are significantly lower for construction jobsites that can enforce some level of control over their environment and personnel.

Building on current trends in industrial automation and physical AI research, we believe there are two components to the high-frequency controller that we must build — a white box real-time collision and safety checking system (safety critic) and a learned action generation policy (adaptive actor). Much of the technology necessary for the safety critic has been matured by the autonomous vehicle industry in the last decade. The technologies necessary for an adaptive actor are seeing increasingly promising results with research in VLA and WAM, but there remains a data gap in training these policies. Our strategy in the meantime is to prioritize construction tasks that can be solved with Model Predictive Control and other optimization methods until we have sufficient real-world data to close the sim-to-real gap and train generalist policies across tasks.

Layer 2 — OperatorAdvanced reasoning AI on the edge

Layer 1 will handle short horizon task execution in real-time but an additional reasoning layer is necessary to prioritize, schedule, and adjust tasks during execution. Such a system would need to not only understand the state of the robot (Layers 0 and 1) and the world around it, but also interact with people and other robots to coordinate task execution. We call this layer the AI Operator — similar to how equipment operators today manage tasks completed by machinery. Our assessment of the state of the art is that modern models like Claude Opus 5, GPT-5.6, and Kimi K3 are likely already sufficient for this layer. We believe this is a matter of distilling a small or medium scale model sized for edge computing, and building the harness and tooling to empower modern language models to manage physical assets.

Layer 3 — OrchestratorCity-scale construction coordination AI agents

Construction jobsites require efficient communication and coordination between workers and crews for smooth operation, so much so that up to a quarter of the personnel on a jobsite are supervisors, project managers, and other supporting staff.^[7] This support crew handles planning and logistics, resource allocation, weather and other environmental monitoring, tool and equipment maintenance, medical support, and engineering support among other tasks. Communication between Mars and Earth has a delay of up to 22 minutes one-way which is a crippling handicap for many of these tasks. Building even a modest 100MW scale solar farm with current technologies would require dozens of staff available on or near Mars to quickly respond to changing conditions. We aim to at least 10x the productivity of supporting crew members with AI agents and tools to enable city building. Similar to AI Operators who will be deployed on the robots, we believe modern reasoning models are sufficient or nearly sufficient for this AI Orchestrator role given appropriate harness and tooling.

Summary

We are living through the convergence of four previously unrelated macro trends:

1. Construction: The US construction industry is facing a massive labor shortage, with estimated 450,000+ worker shortage today^[8]
2. Reindustrialization: The US is reindustrializing and accelerating AI development – $700 billion is being invested in building energy, compute, and industrial capacity^[9]
3. Robotics: The robotics industry is rapidly converging on the capabilities necessary to enable autonomous construction across all layers of the technology stack both hardware and software
4. Aerospace: The aerospace industry is rapidly building the capabilities necessary to send thousands of tons of payload to Mars to begin city scale construction

We’re now entering the window where we can build construction robotics that will not only enable abundance on Earth but also radically improve the chances of building a self-sustaining city on Mars in our lifetime. We believe all four layers of the technology stack are necessary and further believe that tight integration across them is critical. To enable this mission, we have attracted senior talent from leading companies including SpaceX, NASA, JPL, Tesla, Google, Amazon, Zoox, Nuro, and Autodesk. If this is a mission that’s exciting to you, please reach out to the Cosmic Robotics team and apply at the link below.