Feagine Robotics Unveils Fi0 AI for Robot Hardware Adaptability

Feagine Robotics, an organisation, has unveiled Fi0, a cross-embodiment foundation model. This system aims to allow robots to retain task knowledge across machines with varying physical structures. The company also presented three new tendon-driven soft manipulators designed to test this concept, addressing robotics diversification beyond standard industrial applications.

Credit: Feagine

Existing artificial intelligence systems controlling robots typically depend on their specific hardware. Altering a robot's physical form often necessitates significant retraining of its intelligence. Feagine Robotics developed Fi0 to address this. The model processes a robot's physical structure and current state as integral components for generating actions. This approach maintains task consistency, even with physical execution changes, reducing the retraining burden.

Feagine introduced three soft manipulators, A01, A02, and A03, alongside the Fi0 model. These devices vary in lengths, segment counts, and degrees of freedom, specifically for the systematic exploration of cross-embodiment learning. Research suggests exposing models to diverse robot bodies can enhance their ability to adapt to unfamiliar machines. And the company states that Fi0 minimises retraining when robots encounter new tasks.

The A01 manipulator has one flexible segment and two degrees of freedom, weighing 750 grams and carrying a 200-gram payload. The A02 model incorporates two segments, four degrees of freedom, and carries a 400-gram payload. The A03 manipulator, the largest, has three segments, 6+1 degrees of freedom, an arm length of 50 centimetres, and a 600-gram payload capacity. These distinct forms serve as varied test platforms.

According to Feagine Robotics, if Fi0 encounters a task outside its established capabilities, a human operator can provide a single demonstration. This acts as contextual information during inference, avoiding a new training cycle to update system parameters. For instance, a person demonstrating object movement allows the model to extract the task's objects, action sequence, critical contact events, and desired final state. The robot then adapts execution to its own body.

The organisation specifically employs soft, tendon-driven manipulators for these explorations. Traditional robotic arms are mechanically simpler, possessing fixed joints with predictable positions and ranges of motion. Soft manipulators, conversely, bend continuously, alter shape, and interact compliantly with surroundings. But this flexibility makes a robot's physical configuration a critical aspect for the artificial intelligence to comprehend.

The Fi0 architecture integrates information concerning the robot's morphology, sensing mechanisms, actuation methods, and its current state. Feagine refers to this internal representation as an Embodiment Graph. The model also includes components intended to understand the physical environment, interpret human demonstrations, and predict what could happen after different actions. This combination of attributes is what the company terms soft embodied intelligence.

The broader implication suggests robotics may not standardise around a single, universal machine. Diverse environments could instead favour highly specialised robotic bodies. Rigid arms might suit precision manufacturing, soft manipulators could handle delicate interactions. So if the intelligence layer demonstrates portability across these different hardware platforms, the onus of generality shifts away from the physical robot itself.

Humanoid robots have previously garnered interest due to human-centric environment design. General-purpose intelligence, however, does not inherently demand a humanoid form. So Feagine's approach questions whether a singular intelligence can operate numerous specialised bodies, rather than seeking one body capable of all tasks. This signifies a distinct robotics research trajectory.

Fi0 remains an early-generation system. Its full capabilities and assertions require further validation across diverse hardware, tasks, and real-world conditions. And the concept addresses a recognised obstacle in embodied artificial intelligence, ensuring accumulated intelligence does not become redundant with hardware changes. This could lead to a family of specialised machines sharing a growing intelligence.

• Feagine Robotics introduced Fi0, a cross-embodiment foundation model.

• Fi0 aims to allow robots to retain task knowledge across different physical structures.

• The system uses single human demonstrations for new tasks, providing context instead of retraining.

• Soft, tendon-driven manipulators are used for testing, integrating physical configuration into the AI problem.

• This approach suggests a future of specialised robots sharing a common intelligence layer, moving beyond universal designs. Source: interestingengineering