Home›Companies›Sensei

Robotic Training Data at Scale

Sensei helps robotics companies scale and outsource their training data collection. Our hardware platform enables the collection of human-demonstration data at a tenth of the cost and twice the speed of current teleop approaches. Our software platform acts like Scale AI for robotics data: a large network of paid human operators use our low-cost collection platform to fulfill data-generation requests.
Sensei
Founded:2024
Team Size:2
Location:San Francisco
Group Partner:Jared Friedman

Active Founders

Anubhav Guha, Founder

CEO at Sensei - building the marketplace for robotic training data. Previously MIT Robotics PhD (dropped out) and Chief Engineer & Program Manager of Aurora Flight Science/Boeing's DARPA AlphaDogfight team. I've spent my whole career in robotics, and my goal is to help make AI robotics a reality.
Anubhav Guha
Anubhav Guha
Sensei

John Piotti, Founder

CTO at Sensei - scaling robotic training data collection to be 2x faster and 10x cheaper. Previously MIT '19, Robotics Engineer at Aurora Flight Science/Boeing, and full stack web development at UtilityAPI. Also building a house in my free time.
John Piotti
John Piotti
Sensei

Company Launches

tl;dr: Sensei is the Scale AI for robotics training data. Our platform collects human demonstration data at a tenth of the cost and twice the speed of current approaches.

Our Team

We are MIT engineers who have been friends since undergrad. We worked together at Aurora Flight Sciences, where we ran a DARPA-funded program to develop AI algorithms for autonomous fighter jet combat. John went on to spearhead the reinforcement learning efforts at Aurora, while Anubhav returned to MIT for a PhD focused on robotics, control theory, and machine learning (but dropped out to found Sensei!).

We have spent our careers working at the frontiers of robotics and AI. With Sensei, we plan to empower the expansion of those frontiers.

The Problem

Data scarcity is one of the biggest challenges in developing robotic artificial intelligence.

Innovators in the space are looking to apply the learnings from the Foundation Model Era: that large models trained on huge datasets are incredibly powerful. As researchers, engineers, and a rapidly growing number of commercial efforts race to develop large models for robotic intelligence, it’s become increasingly clear that the availability of sufficient quantities of quality data is the most significant bottleneck.

Current solutions are slow, expensive, and fundamentally not scalable. Human demonstrations are the gold standard for training data. Robotics companies currently employ fleets of 5-20 data-collectors. These in-person contractors teleoperate a robot to perform hundreds of demonstrations a day - with tasks ranging from clothes folding, to bin sorting, to dishwasher loading. This is not the right long-term solution:

  • Large-scale collection is prohibitively expensive: a teleop setup requires a physical robot, often with a cost of $40,000+ per operational platform.
  • Low-quality data: restriction to labs/office spaces makes it hard to collect varied and realistic demonstrations
  • Slow data collection rates: non-intuitive interfaces, cumbersome equipment, and faulty hardware lead to slow demonstrations on platforms that break often.

These drawbacks make it impossible to scale quality training data collection to the quantities needed for robotics.

Our Solution

We design and manufacture low-cost and easy-to-use devices for collecting human demonstration data. Our platform costs less than $300 and can be used by anyone to collect high-quality training data. As seen in a clothes folding task that demonstrates our research prototype, the operator is equipped with a sensorized exoskeleton arm that closely matches the natural human range of motion. The intuitive design, coupled with a suite of angle, vision, and inertial sensors, enables the rapid collection of highly accurate visuo-spatial state information.

Our main advantage is that an operator can easily set up and equip the platform, leading to an unprecedented ability to generate quality data at scale. Demonstrations can be performed for a broad range of common tasks, in a nearly infinite number of diverse settings, in as many different ways as human behavior is varied.

In order to maximally utilize a fleet of our data-collect platforms, we are building out and operating a network of Senseis— contractors that have been trained to collect high-quality training data using our devices. Our Senseis receive task descriptions from robotics researchers and engineers, and are then paid to collect demonstrations that fulfill the request.

Our combined hardware + software stack represents the first truly scalable solution to generating training data for AI robotics.

Our Goal

Within the next five years, we envision tens of thousands of operators, equipped with increasingly sophisticated, portable, and effective data-collection arms, exoskeletons, headwear, and tools. These Senseis will be located all over the world, come from and live in a variety of environments, and all have different takes on what it means to perform tasks “like a human.” This is the best way to power data collection for improved robots, and it is the future we’re building.

Ask

If you’re interested in solving the data scarcity problem in robotics - either for yourself, your company, or your customers - we’d love to hear from you. You can reach us at founders@senseirobotics.com