VLA Decoding Notes (2): The Action Engine — Diffusion Policy, Flow Matching, Low Latency

1. Why Discrete Tokens Fall Short

As noted, quantizing continuous actions into autoregressive tokens truncates precision and stutters. Robots need high-frequency, smooth, perturbation-sensitive continuous control — exactly what diffusion / flow matching deliver.

2. Diffusion Policy: Denoise Actions Like an Image

Diffusion Policy (Chi et al., 2023) borrows image-generation diffusion:

• Training adds noise to action sequences; the network learns “recover action from noise”;
• Inference starts from random noise and denoises over steps into a multimodal, robust trajectory.

It naturally supports multi-modal distributions and resists noise in demonstration data.

3. Flow Matching: Faster and More Stable Than Diffusion

Flow Matching does not model a complex probability path; it learns the mapping of a probability flow, pushing a simple distribution straight toward the target action distribution.

• Simpler training objective, steadier gradients;
• Far fewer inference sampling steps than DDPM (tens to hundreds) — down to ~5 steps;
• π0 and Xiaomi both use Flow Matching as the action expert’s core.

The “flow matching” we covered in the speculative-decoding notes here lands as physical action, not text tokens — the same math tool, reused across modalities.

4. Action Chunking: Predict a Whole Block at Once

Deciding every millisecond is slow and jittery. Action Chunking (from ACT / π0) lets the model predict a short block of actions at once:

• Lower decision frequency, smoother motion;
• Intra-block autoregressive/diffusion generation, inter-block prefix from history;
• π0 uses Flow Matching + 50Hz chunks — the “smooth + high-frequency” recipe.

5. Xiaomi’s Play: MoT Loosely-Coupled + Async Inference

Xiaomi-Robotics-0 (open-sourced 2026-02, 4.7B) is a low-latency VLA exemplar:

• MoT: VLM “brain” understands; 16-layer DiT (Diffusion Transformer) “cerebellum” generates action blocks. Loosely coupled via KV Cache — brain output feeds the cerebellum, no recompute.
• Flow Matching training: sampling steps cut from DDPM’s tens to 5.
• Async inference: model inference and robot execution are decoupled — latency no longer stalls real-machine continuity, killing “action断层” at the mechanism level.
• Clean Action Prefix + Λ-shape attention: prior-step actions as input keep temporal continuity; a special mask makes the model watch current visual feedback, reacting sharply to sudden changes.

Result: 80ms latency, 30Hz control, real-time on RTX 4090, SOTA on LIBERO / CALVIN / SimplerEnv.

6. Bridge

The action-generation mechanism is clear. Next: how the π family evolved with it — from π0’s Flow Matching 50Hz to π0.7’s memory and compositional generalization.