Our vision is to revolutionize robotics by eliminating the slow and costly nature of traditional robot development. We aim to create a platform that generates robot designs alongside their optimal control policies, streamlining the process and opening up new possibilities.
"I think one could make progress in robotics today. … Build many thousands, tens of thousands, hundreds of thousands of robots and somehow collect data from them and find a gradual path where the robots are doing something slightly more useful."
Chief Scientist of OpenAI
The time is now
Grout & tile cleaning
Aquatic weed removal
Ship hull cleaning
Cleaning solar panels
Having taken the initial step by developing our first grouting robot using conventional methods, we now move to the next phase. This involves training a control policy with reinforcement learning in simulation and running it on the Grouting Robot V4, demonstrating the capability for multi-hour task execution.
As we progress, we'll identify the next task for which we want to create a robot. Utilizing our platform, we'll generate multiple robot designs for the new task, build and test them in parallel, and ultimately bring one to revenue, showcasing the potential of our groundbreaking approach.
Our focus lies in a subset of robotics tasks that do not involve external agents. We'll start with tasks similar to grouting and gradually move to any tasks without external factors, tackling challenges of increasing complexity while refining our platform and expanding its capabilities.
Upon reaching a suitable level of development, we'll open the platform to the public, transforming it into the new CAD for robotics innovation.
Accelerating the robot design process is achieved by shifting the burden from developing software controls to creating high-fidelity simulations. The term "high-fidelity" holds different meanings for various tasks. For certain tasks, a highly detailed and realistic simulation is essential for accurately assessing robot performance. On the other hand, some tasks might only require basic representations of the environment and interactions.
We envision training a foundational model capable of fine-tuning for specific robotics tasks, which would then be used to predict the immediate outcomes of events based on intricate knowledge of the possibilities. Full future distribution. For example, the model could anticipate the effectiveness of grout application, the likelihood of a nail breaking under the robot's force, potential damage to objects upon contact, or the breaking point of aquatic weeds during removal.
To ensure the model's predictions closely align with reality, it is crucial to collect data from a large fleet of robots that interact with the diverse, real-world industrial environment, as opposed to being confined to predefined spaces like car manufacturing robots. While companies like Google and Facebook AI have developed unbounded robots, they do not have thousands-strong fleets for large-scale data collection and evaluation.