ONE LOOP
A VLA is easiest to understand as a repeated perception-to-action loop.
1 · OBSERVATION
The policy first receives a view of the current embodied situation.
Visual input may come from one camera, multiple external cameras, wrist cameras or video. Many VLA systems also use numerical robot state such as joint positions or gripper state. Language conditions the behavior: pick an object, open a drawer, sort parts or complete a longer task.
2 · MULTIMODAL ENCODING
A vision-language backbone turns pixels and words into shared context.
OpenVLA uses visual encoders plus a language-model backbone. SmolVLA uses SmolVLM2. GR00T uses an NVIDIA Eagle VLM. The exact architecture differs, but the recurring idea is to reuse visual-language pretraining so the robot policy starts with semantic representations of objects, scenes and instructions rather than learning every concept only from robot trajectories.
3 · EMBODIMENT STATE
Robot state makes the same instruction mean something different on different bodies.
Two robots may see the same cup and receive the same instruction but have different reachable workspaces, grippers, joint limits and control interfaces. Proprioception and embodiment-specific state ground the policy in the body that must actually act.
4 · ACTION GENERATION
The action head converts multimodal context into a robot action space.
Action tokens
Continuous actions are quantized or compressed into tokens and predicted autoregressively.
Action expert
A dedicated policy head predicts continuous robot actions conditioned on the multimodal representation.
Action chunks
The policy predicts several future control steps together rather than only one action per expensive model call.
5 · EXECUTION
Model output still has to cross a robot control boundary.
The predicted action representation must be mapped to the target robot's control interface. Depending on the system, the VLA may output end-effector deltas, joint-space targets, gripper values or another policy-specific vector. Controllers, calibration, timing and actuator interfaces then turn those targets into physical motion.
6 · CLOSED LOOP
The next observation is part of the policy, not an afterthought.
Physical action changes the scene. Objects move, contacts fail, humans intervene and the robot itself changes configuration. Re-observation closes the loop and allows the VLA policy to correct its next action rather than relying on an open-loop plan that assumes the world stayed unchanged.
TRAINING
VLA training binds visual-language representations to robot trajectories.
Training examples usually connect observations and instructions with recorded robot actions. Demonstrations may come from teleoperation, autonomous collection, simulation, synthetic generation or large cross-embodiment datasets. The model learns which action representation is compatible with the multimodal context and task.
OPENRGD VIEW
A learned policy can be powerful without owning every downstream responsibility.
OpenRGD treats the VLA as a cognition provider inside a wider embodied architecture. That preserves a place for explicit body semantics, capability grounding, safety gates and hardware-specific adapters when a deployment needs those boundaries.