Intelligence Factory Builds Multimodal Data Infrastructure to Push Robotic Manipulation to 99% Reliability

While general-purpose robotics models often struggle with success rates around 40% to 50% outside structured environments, Intelligence Factory is tackling the reliability bottleneck in physical automation. Founded in 2026 and backed by Y Combinator's P26 batch, the five-person San Francisco company develops foundational manipulation models designed to take robot performance to 99% success rates across commercial environments including grocery fulfillment, warehouses, data centers, and electronics assembly.

Rather than manufacturing custom robot arms or humanoids, Intelligence Factory focuses on intelligence and data collection. The team operates specialized facilities where human data collectors wear custom sensorized gloves to capture multimodal demonstration data—combining vision, motor action, and force sensing. This demonstration data is used to post-train foundation architectures into task-specific, hardware-agnostic control policies that adapt to new robot arms in 20 to 25 minutes of training.

Founded by Manipulation and Controls Engineers

The company was co-founded by CEO Yash Sinha and CTO Jalaj Shukla. Yash Sinha studied robotics at ETH Zurich, previously founded Kaio Labs, and worked as an F1 simulation engineer. Jalaj Shukla holds a Master's in Robotics and Mechatronics from the University of Pennsylvania's GRASP Lab and served as a research fellow at India's Defence Research and Development Organisation (DRDO). Prior to co-founding Intelligence Factory, Shukla worked on manipulation, vision-language-action (VLA) models, and controls at Dimensional, Blue Sky Robotics, Applied Materials, and Ansys.

Speaking on The Parameter, Sinha explained that current foundation models fall short in real-world deployments because internet-scale data does not exist for physical interaction. "All the whole robotics community is pushing towards these bigger better pre-trained models that are highly generalizable but eventually their success rates are between 40 to 50%," Sinha said. "Robotics cannot be a commercially valuable proposition if it doesn't work at 99%."

In an interview on Code Story, Sinha noted that the company deliberated whether to monetize its dataset directly or build proprietary deployment capabilities. The team chose deployments to establish long-term defensibility, building models capable of fine-tuning across third-party robot hardware within 25 minutes.

Hardware-Agnostic Models and Data Collection Hardware

To supply its training pipeline, Intelligence Factory relies on custom wearable capture hardware built in-house. The company operates two main hubs: software and model engineering in New York, and hardware design in Pune, India.

To lead its model architecture and training pipeline, the company is recruiting a Founding Engineer: Manipulation in New York ($80,000 – $200,000 / year). The role owns the end-to-end model pipeline, including curation, training, evaluation, and deployment of foundation models for manipulation robots, selecting model architectures and setting technical direction.

On the hardware side, the team is hiring an Electronics Engineer (India) in Pune, Maharashtra, India (₹1,200,000 – ₹1,500,000 / year). This role owns the electronics and firmware for custom data collection hardware, including wearable instruments used to collect tactile and force data.

In addition, Intelligence Factory is bringing on a Robotics Intern in New York ($5,000 – $7,500 / month) to evaluate manipulation models on real physical robots in the lab, build simulation evaluations, collect demonstration data using capture hardware, and maintain robot arms and sensors.

Open Roles at Intelligence Factory

Below are the open positions listed at Intelligence Factory:

See all open roles at Intelligence Factory