Every Vision-Language-Action (VLA) model makes a bet that most papers bury in a methods section: how do you turn a transformer’s output into a robot command? The choice looks like an implementation detail. It isn’t. Action representation — whether you discretize, diffuse, or use flow matching — determines inference latency, action precision, multi-modality handling, and how efficiently the model uses its context window. As VLAs move from lab demos to real deployments, this is the design decision that separates systems that work from systems that almost work.

1️⃣ The discretization path, and its ceiling. RT-2 and OpenVLA took the obvious route: bin each continuous action dimension into 256 buckets and append the resulting tokens to the model’s vocabulary. The appeal is clean — you get a single unified autoregressive model, no extra architectural components, and the full weight of pretrained language representations. The cost is threefold. Discretization introduces quantization error that compounds over a 6-DOF arm into millimeter-level jitter on fine tasks. Autoregressive decoding generates one token per forward pass, so a 7-DOF action at 10 Hz requires 70 forward passes per second — a latency budget that kills real-time control unless you run on expensive accelerators. And because each dimension is tokenized independently, the model has no inductive bias toward the temporal and spatial correlations that define smooth motion. OpenVLA’s fine-tuning results on BridgeV2 are impressive, but the hard manipulation tasks — in-hand reorientation, compliant insertion — remain systematically worse than methods with continuous heads.

2️⃣ Diffusion heads: precision at a price. Physical Intelligence’s π₀ made a different bet. Its architecture keeps a pretrained vision-language backbone frozen (or lightly adapted) and bolts on a diffusion-based action head that iteratively denoises a Gaussian prior into a clean action trajectory. The payoff is real: diffusion naturally handles multi-modal action distributions (the classic “move left or right around the obstacle” ambiguity), produces smooth trajectories, and can represent high-precision continuous values without discretization error. On the DROID benchmark and in π₀’s own zero-shot transfer evaluations, diffusion heads consistently outperform discrete tokenizers on contact-rich tasks. The cost is inference: even with DDIM acceleration, 10–25 denoising steps per action chunk adds 50–100 ms of latency that makes closed-loop reactive control difficult. The workaround — action chunking, predicting a short horizon of future actions at once — trades off responsiveness for throughput. It works, but it’s a patch, not a fix.

3️⃣ Flow matching closes the gap. The emerging consensus is continuous normalizing flows via flow matching, which train a vector field that transports samples from noise to action in one or very few integration steps. Where diffusion requires iterative refinement, flow matching learns a straighter probability path, enabling near-single-step inference without sacrificing distributional expressiveness. π₀.5 and several concurrent preprints from CMU, Berkeley, and ETH Zürich have shown that flow-matching action heads match diffusion quality on multi-modal benchmarks while cutting inference time by 4–8×. The framing matters: flow matching isn’t just “faster diffusion” — the straighter transport paths also improve training stability and data efficiency, which matters enormously when robot datasets are measured in tens of thousands of demos rather than billions of tokens.

4️⃣ What the architecture choice really encodes. The deeper issue is that action representation choice is a statement about what kind of policy you believe you’re learning. Discrete tokens assume actions are compositional symbols; diffusion assumes they are samples from a complex learned distribution; flow matching assumes a smooth, learnable transport map exists between noise and behavior. Each assumption breaks differently: discretization fails at precision, diffusion fails at latency, flow matching is still being stress-tested on highly dynamic tasks like bipedal recovery where the action distribution may be genuinely discontinuous. No single answer has won.

The field is converging toward hybrid architectures — a large pretrained backbone for perception and language grounding, plus a lightweight continuous head (flow matching preferred) for action generation — with action chunking as an orthogonal knob for latency/reactivity tradeoff. The next frontier is making the head architecture-aware of physical priors: contact constraints, joint limits, dynamics. The representations that win will be the ones that speak the language of physics, not just the language of tokens.