Cortex Robot, Inc.

Cortex Robot, Inc., which trades as Cortex AI, is a San Francisco robotics data startup founded in 2025 and backed by Y Combinator as part of its Fall 2025 batch, with Garry Tan as its primary YC partner. Its YC profile lists Lucas Ngoo, previously co-founder and CTO of the Singapore marketplace Carousell, as founder and CEO, and a team of three at the time of listing; the company's current job openings are based in San Francisco, Singapore and Johor Bahru, Malaysia, covering robot operators, operator managers, a robotics hardware engineer and an internal-tooling software engineer. A third-party funding database reports one round totalling USD 6 million.

Cortex's stated goal is to build a large and diverse real-world, real-workplace and industry-scale dataset of egocentric human video and robot data for embodied AI. It sells three services to frontier labs and robot developers. The first is egocentric data for robotics and world models, captured in real workplaces with every frame annotated for hand and body pose, depth and subtasks so that first-person video can be used directly for robot learning; the company recruits contributors through an Operator Network and, according to YC, runs the Cortex Marketplace, which pays workplaces to host data-collection sessions. The second is robot trajectory data for fine-tuning, collected on robots placed in real workplaces through industry partners; published sample tasks include making instant coffee and bagging groceries with recovery behaviour, recorded as mobile bimanual robot data from real deployments and simulation. The third is human-in-the-loop rollouts and evaluations, in which people supervise real-world deployments and capture recovery data for edge cases to form a continuous data flywheel.

In September 2026 Cortex published an evaluation of large language models controlling a bimanual YAM BOX robot with an overhead camera and two wrist cameras, extending the real-world evaluation protocol of the MolmoAct 2 paper. The models attempted to load a sandwich into a microwave and close the door, and to attach a 5 mm tip to a pipette, with up to ten minutes per attempt and partial credit for completed steps. In October 2026 the company announced Cortex Harness, a platform for running large language models and vision-language-action models, including open-weight models, on real robots under remote human guidance. Operators choose a model, describe the task and set reasoning effort, then watch the camera feeds and the model's reasoning, steer it with additional prompts or take direct remote control, before handing control back. Interventions are logged as recovery data for training; the launch demonstration showed an operator in Singapore guiding a robot in San Francisco to insert a battery into a slot. Cortex Harness is offered through a waitlist and beta programme with usage credits.