While humanoids race to the finish line with new companies releasing their prototypes in quick succession, one player seems to be on cruise mode with its vision of commercialising humanoids – Figure AI.
Backed by big players like Microsoft and OpenAI, Brett Adcock, the founder and CEO of Figure AI, recently spoke about the compute and inference that goes into Figure humanoids, and was confident about never facing a shortage for scaling.
“Microsoft has given us as many H100 GPUs as we needed. So we’ve scaled up three times [in] 90 days, into large clusters,” said Adcock, adding that they are not bound by compute for training and experiments.
With the infrastructure in place, Figure AI’s path ahead is relatively simpler. The company is concentrating on executing the crucial task of speed and computation for its humanoids.
Horizon Tasks for Speed and Computation
Figure AI is placing significant emphasis on utilising “horizon tasks” in humanoid robots to optimise efficiency and performance. Horizon tasks, which refer to long-duration tasks such as household chores, rely on a balance between onboard and off-board computation.
The company highlights the advantage of leveraging larger external models or big computers to handle complex tasks, while leaving faster, localised actions to the robots themselves.
When tasks such as “putting away the laundry” or “organising toys” arise, the more computationally intensive aspects, like planning and reasoning, can be off-boarded to high-powered external servers. These tasks, categorised as longtime horizon tasks, allow for the use of mega computers that process information within one or two seconds.
Adcock mentioned that smaller tasks could be off-loaded, processed by these mega computers, and the results could be returned within one or two Hertz.
On the other hand, actions that require real-time responses, such as fine motor movements, are handled locally, utilising onboard GPUs that operate at much faster speeds. “We can run that separately at 100 Hertz,” he explains, emphasising that actions like moving an arm or grasping objects must happen instantly without waiting for external input.
Hand-movements have always been a crucial make-or-break criteria for any form of humanoid robots as it involves intrinsic coordinative movements.
The founder of Bengaluru-based robotics company CynLr Robotics, Gokul NA believes that today’s robotics developments, especially robotic arms, are more of ‘record and playback machines’ with sophisticated manipulation. However, they lack in perception.
“In most cases where you want to commercially deploy these robots, you don’t need legs. Wheels are more than enough, but you need more capability with the hands,” he said.
Why the Dual Approach?
The dual approach allows humanoid robots to manage a mix of both highly complex and rapid-response tasks. This method ensures robots remain functional even in environments where internet connectivity may be unstable, as onboard models, though smaller, are becoming increasingly capable.
Adcock goes on to explain the necessity for fast internet access to offload complex tasks to larger, external computers for quicker processing. “If it’s really fast, you can do everything,” but connectivity isn’t always guaranteed.
In situations where the connection is lost, robots must continue working using onboard capabilities. This balance between onboard and offboard processing is critical for real-time actions. Adcock is confident that as AI models improve, with predictions that they will be “smaller and better” by 2028. Robots will increasingly handle more tasks locally, even without a consistent internet.
While Figure AI’s dual approach allows it to function autonomously (though limited) even without connectivity, other humanoid manufacturers also have this capability. Boston Dynamics’ Atlas leverages its onboard sensors and processing to perform complex tasks without relying on constant internet connectivity.
In April this year, the robotics company suspended the development of its hydraulically actuated robot and replaced it with an electric version of Atlas with superior capabilities and dexterity.
Similarly, Tesla’s Optimus is designed to perform autonomously, possibly hinting at its functionalities without connectivity either.
Comparing with Tesla
A notable competitor in the humanoid space for Figure AI is without doubt Tesla. However, Adcock doesn’t focus on comparing themselves to Tesla, as he opines they are on their own vector with a clear roadmap aimed at home integration through workforce applications.
“Tesla overall is doing great. I think their vector of where they’re heading is in the right direction,” he said. However, he disagreed with the notion that companies should skip workforce tasks and head straight to home-based robots. He stressed that robots are not yet robust enough for complex home environments, and focusing on workforce tasks will help improve reliability, reduce costs, and train AI systems more efficiently.
The post Figure AI’s Embodied Humanoids Perform Tasks with Little to No Internet appeared first on AIM.
