Training Data for Figure AI

Figure AI's approach to general-purpose humanoid robotics represents the highest-resource bet in the industry. With $2.6 billion in funding and partnerships with OpenAI, Microsoft, and NVIDIA, Figure has the capital to pursue the data-intensive path to humanoid intelligence. Their custom Helix VLA model — built in collaboration with OpenAI and leveraging OpenAI's language model expertise — follows the paradigm established by RT-2: co-train a vision-language model backbone on both web-scale data and robot demonstrations to produce a model that reasons about novel situations and translates language instructions into physical actions.

The 41-degree-of-freedom action space of Figure 02 creates a fundamental data scaling challenge. A standard 7-DOF robotic arm needs demonstrations covering a 7-dimensional action space. Figure 02's full body — 41 joints including two 16-DOF dexterous hands — requires coverage of a 41-dimensional space. The curse of dimensionality means that achieving equivalent coverage requires exponentially more demonstrations. This is not a theoretical concern — it is the primary bottleneck Figure faces in scaling from impressive laboratory demonstrations to reliable commercial deployment.

The dexterous hands are particularly data-hungry. Each hand has 16 degrees of freedom allowing individuated finger control, precision pinch grasps, power grasps, and tool manipulation. The contact dynamics between robot fingers and real-world objects (rigid, deformable, fragile, slippery) are nearly impossible to simulate faithfully — friction, compliance, and texture vary across millions of object types in ways that no physics engine captures. Real-world manipulation data with force and tactile feedback from diverse objects is irreplaceable for training dexterous hand policies.

Figure's BMW deployment at the Spartanburg, South Carolina manufacturing plant provides initial commercial validation but also exposes the generalization challenge. A policy trained at Spartanburg with specific tooling, part geometries, and environmental conditions must transfer to other manufacturing facilities with different configurations. Each new deployment site requires either massive retraining or initial training data diverse enough to cover the variation across sites. Claru's ability to collect manipulation data across dozens of industrial environments directly addresses this generalization requirement.

The sim-to-real transfer problem is particularly acute for humanoids. Simulated locomotion policies trained in Isaac Gym or MuJoCo consistently fail on real hardware due to ground contact dynamics, actuator backlash, and compliant surface interactions. Real-world walking data collected in actual warehouse and factory environments — with realistic floor textures, obstacles, and human co-workers — provides the distributional coverage that simulation alone cannot deliver.