WHY IT MATTERS
Action representation shapes precision, latency and portability.
A robot ultimately moves in continuous physical space, but the learned policy does not have to represent that motion in one universal way. The model may quantize actions into tokens, directly generate continuous values, denoise a trajectory with diffusion-style objectives or integrate a learned flow field.
MAIN FAMILIES
Four recurring ideas dominate current VLA action design.
Discrete action tokens
Continuous robot actions are quantized into vocabulary-like symbols and predicted autoregressively.
Continuous action prediction
A dedicated action head predicts continuous values rather than reusing a language-token output space.
Diffusion / flow generation
The policy learns to transform noise into a coherent continuous action trajectory conditioned on multimodal context.
Action chunks
Several future actions are predicted together so high-cost inference does not have to run at every low-level control tick.
DISCRETE TOKENS
Tokenization makes robot action look like another language-model vocabulary.
OpenVLA's original design predicts tokenized actions from its language-model backbone and then decodes those tokens into continuous actions. This is architecturally convenient because the same autoregressive machinery used for text can be reused for robot control.
The trade-off is that a continuous control space has to be quantized or compressed. Later work around OpenVLA introduced FAST-style tokenization to compress action chunks into fewer tokens and improve inference speed.
CONTINUOUS ACTIONS
A separate action expert can preserve continuous control structure.
Continuous-action VLAs keep multimodal understanding in the vision-language backbone while delegating robot action generation to a specialized policy head. This avoids forcing every control value into a discrete text-like vocabulary and can better match high-frequency or dexterous control requirements.
FLOW MATCHING
π0 and SmolVLA show why flow-based action generation became important.
π0 trains its action expert with conditional flow matching: the policy learns a vector field that maps noisy action trajectories toward demonstrated robot actions. At inference time, the trajectory is generated by integrating that learned field. SmolVLA also uses a flow-matching action expert conditioned on VLM features.
ACTION CHUNKING
A policy can reason slowly and still emit a faster temporal control sequence.
Predicting a chunk of future actions amortizes expensive multimodal inference across several control steps. It can improve throughput and temporal consistency, but it also introduces a horizon choice: longer chunks reduce inference frequency while making the policy less immediately reactive to newly observed changes.
LATENCY
Model frequency and actuator frequency are not the same number.
A VLA may run at tens of inferences per second while lower-level controllers operate much faster. Asynchronous inference, chunked actions and downstream interpolation are different strategies for bridging that gap. SmolVLA explicitly demonstrates asynchronous inference that decouples action execution from the next model prediction.
EMBODIMENT
The output vector only has meaning relative to a robot-specific action space.
A six-dimensional end-effector delta, a joint-position vector and a whole-body humanoid action are not interchangeable even when all are represented as continuous numbers. Action normalization, joint ordering, reference frames, gripper semantics and control mode remain embodiment-specific contracts.
OPENRGD BOUNDARY
ActionIntent should describe intended action semantics without pretending every robot shares one control vector.
OpenRGD's candidate architecture separates a model-level action intent from hardware-specific execution. That gives an embodied runtime space to resolve capabilities, frames, constraints and Body Adapter mappings before an action reaches device-level control.