VLA
Vision-Language-Action model. A neural network that takes images and natural-language instructions as input and outputs motor commands. The successor to VLMs for robotics, where the action head closes the loop from input to motion.
Also known as: vision-language-action, robotic foundation model
A VLA extends the vision-language model architecture with an action head. The encoder ingests camera frames (sometimes lidar, depth or proprioception) alongside a natural-language goal; the decoder emits tokens that translate into low-level motor commands (joint torques, end-effector poses, gripper states, wheel velocities). Everything turns on the training data: pairs of (observation, instruction, expert action trajectory) collected from teleoperation, simulation, or YouTube-style human-demonstration corpora. Google’s RT-2, DeepMind’s RT-X and Physical Intelligence’s π0 are the canonical published examples; many humanoid OEMs are training their own.
The reason VLAs matter to DeAI infrastructure is that they pull large-model inference onto the robot itself or to the very edge. A humanoid acting on a one-hertz reasoning loop cannot tolerate round-trip latency to a hyperscaler GPU. The runtime layer that ships on the robot (Openmind’s OM1, Nvidia Isaac, Physical Intelligence’s stack) handles the local inference, sensor fusion and action dispatch, with the cloud reserved for higher-level planning or retraining. That architecture is where “local AI” stops being an enthusiast position and becomes a hardware requirement.
Verifying a VLA’s behaviour is the open problem. A misbehaving language model produces an embarrassing paragraph; a misbehaving VLA can break things, hurt people or steal funds it has been given custody of. On-chain identity standards like ERC-7777, attestation schemes like the FABRIC verification layer, and the broader work on Proof-of-Robotic-Work emissions are all attempting to bind a robot’s actions to a verifiable identity so the network can reward correct behaviour and slash incorrect behaviour. The cryptographic side of that work is ahead of the regulatory side, which is ahead of any consumer’s ability to evaluate a VLA before they put it in their kitchen.