Redefining Robotic Intelligence

Designing intelligence for the real world

Our team focuses on deploying robotic systems into the real world. We build general purpose foundation models that can adapt to the variability of commercial and industrial environments.

A path to physical AGI

FutureVision brings the capability to handle real world industrial tasks autonomously. This unlocks generalization that conventional VLA pipelines struggle to achieve. We call this the Direct Video Action model. And it’s why Rhoda works not just in labs but in factories, warehouses, and real production environments.

PRE-TRAINING

1+ Years

of web-scale video data

POST-TRAINING

1-10 Hours

trajectory data

We first pre-train our model with over a million videos, giving it a strong prior on motion, physics, and dynamics. We then post-train on action data collected on the robot to teach the model to learn specific tasks. This combination of pre- and post-training creates generalist robot policies that can work and adapt in dynamic environments.

We make robots smarter

We explore a new paradigm for scaling robot intelligence with web-scale video pre-training. We’ve published an in-depth research blog to discuss the novel architectural designs and the benefits of such an approach.

View Research Blog

Returns processing

This is an end-to-end returns processing task for a customer in the logistics industry. The task contains ambiguity—visually similar states can correspond to very different points in the pipeline. Rather than relying on the current frame alone the model maintains memory, in the form of a long history of frames.

Bearing decanting

This task comes from an automotive assembly line. The customer initially believed it could not be automated because of several challenges: each box weighs 10 kg, the lifting strap can easily tear, removing the tab requires precise control, and the transparent plastic bag is difficult for a robot to grasp.

Contico breakdown

Contico are 50-pound, heavy-duty boxes that are ubiquitous in manufacturing for transporting materials between facilities. After use, they must be manually cleared of debris of random sizes and type, unlatched, and collapsed for return or storage. This task is difficult to automate due to the large box size and variability in debris.

Human demo following

Long-context enables in-context learning, allowing us to inject human demonstrations into our robot's context window. This enables our robot to perform tasks single-shot, without retraining. Here we demonstrate single-shot pick and place and single-shot drawing from human demonstrations.

Commercial
Applications

We work with a variety of customers across verticals in automotive, manufacturing, logistics, and ecommerce. If you’re interested in working with us at your facility, reach out here.