Foundation Future Industries, also known as Foundation Robotics, has entered a strategic collaboration with AMD to train and operate its Phantom humanoid robots using AMD computing technologies.
The companies are developing an embedded computing architecture for Foundation’s production robots using AMD Kria AI system-on-modules powered by AMD Ryzen AI Embedded X100 Series processors.
Foundation has validated its Phantom MK-1 humanoid robot on AMD compute with full locomotion. Its next-generation Phantom MK-2 platform will be built natively around AMD processors, with full locomotion. field-programmable gate arrays and other embedded technologies.
The collaboration covers the development of an X100 Series-powered embedded “brain” for the robot, along with AMD FPGA-based systems supporting Phantom’s hands, sensing and real-time control.
FPGAs are programmable semiconductor devices that can be configured to execute specialized tasks with predictable timing. In robotics, they can help coordinate motors, process sensor inputs and perform control operations with low latency.
Foundation said its custom robotic hands will use AMD FPGAs to coordinate 23 degrees of freedom while processing high-frequency tactile feedback through a single deterministic control loop.
Degrees of freedom refer to the independent movements a robotic system can perform. A hand with 23 degrees of freedom can coordinate numerous joints and motions required for gripping, repositioning and manipulating tools or objects.
The companies are showcasing the collaboration during AMD Advancing AI 2026 at the Moscone Center in San Francisco. Phantom MK-1 is performing live demonstrations in the event’s Demo Showcase on July 22 and July 23.
Foundation CEO Sankaet Pathak also participated in a conference panel focused on developing AI-powered robots for real-world operating environments.
Foundation is developing humanoid robots for heavy industrial and defense applications, including manufacturing, construction, port operations, border control and other environments where equipment may encounter water, dust, vibration and extended work schedules.
The company said its robots currently operate 24 hours a day, five days a week in customer deployments. Foundation attributed its systems with contributing to the production of more than 24,000 vehicles during 2025.
Foundation also reported $100 million in contracted annual recurring revenue. The company did not disclose the customers, duration, deployment schedules or remaining conditions associated with those contracts.
The upcoming Phantom MK-2 is being engineered for demanding industrial conditions. Foundation said the robot is designed to withstand shocks and vibrations of up to 100G and will carry an IP67 rating.
An IP67 rating indicates that an enclosure is designed to prevent dust from entering and can withstand temporary immersion in water under defined conditions.
The robot will also incorporate 360-degree vision to improve environmental awareness while working around equipment, materials and people.
Foundation believes this durability will allow Phantom MK-2 to operate through complete industrial shifts in locations that may be unsuitable for less rugged humanoid systems.
The company’s work with AMD is also intended to address the gap between simulated robot performance and behavior in real environments.
Robotics developers frequently train AI models and movement policies in simulation before deploying them onto physical machines. Performance can decline when the robot encounters sensor noise, timing differences, unpredictable objects or physical conditions that were not accurately represented during training.
Foundation said deterministic control, low-latency processing and the ability to analyze large volumes of sensor information in real time are important for reducing that gap.
The company claimed that its AMD-powered architecture delivered approximately 2.5 times faster CPU inference and three times faster training than the NVIDIA platform used for its comparison, while producing equivalent model quality. Foundation did not disclose the specific NVIDIA hardware, models, workloads or testing methodology used for the benchmark.
Inference is the process through which a trained AI model interprets new information and makes decisions. For a humanoid robot, inference can support functions such as recognizing an object, determining how to grasp it or adjusting movement in response to changing surroundings.
Faster training can help engineers develop and improve robotic behaviors more quickly, while faster inference can reduce the delay between receiving sensor data and performing a physical action.
AMD said the collaboration combines its AI acceleration, CPU, graphics and adaptive computing technologies to support the broader robotics pipeline.
That pipeline includes perception, reasoning, autonomous decision-making, tactile sensing, hand manipulation and the real-time control of motors and other components.
Foundation was founded in May 2024 and is headquartered in San Francisco. The company describes its robots as dual-use systems because they are intended for both commercial industrial environments and defense-related applications.
The collaboration gives AMD another opportunity to apply its embedded computing technologies beyond traditional computers and data centers. For Foundation, adopting an integrated AMD architecture could provide greater control over the processors, programmable hardware and software used across its robotic system.
KEY QUOTES:
“A foundational problem in robotics is closing the gap between sim and real. That comes down to not just models, but determinism, latency, and processing rich sensor data in real time.”
“With approximately 2.5 times faster inference, equal model quality and three times faster training speed compared to NVIDIA, the AMD-powered platform gives us the performance we need to put robots to work at scale.”
Sankaet Pathak, Co-Founder and CEO of Foundation Robotics
“With AMD FPGAs, Foundation’s custom hands will actuate and coordinate all 23 degrees of freedom while fusing high-frequency tactile feedback in a single deterministic control loop, giving Phantom human-like manipulation performance.”
Andrea Esposito, Hand Development Lead at Foundation Robotics
“Advanced humanoid robots require tightly coordinated compute across perception, reasoning and real-time control.”
“Our collaboration with Foundation brings together AMD AI acceleration, CPU, graphics and adaptive computing technologies to support the complete robotics pipeline, from high-frequency sensing and manipulation to autonomous decision-making.”
Salil Raje, Senior Vice President and General Manager of AMD Embedded
