RobotGuide
Brain and cerebellum/Deployment and Inference

Foundation-model teardown · Deployment and Inference · 20 min

Deployment and Inference: landing a 1–10Hz brain

The latency budget table, frequency decoupling, five optimization levers, and the safety reflex layer — every line of the math that turns '50Hz in the paper' into 'runs on the production line'.

20 minreal anchors, sourcedchecked 2026-11

What it really is, once the buzzwords are off

The problem at deployment is never 'the model isn't good enough' but 'the model's tempo doesn't match the machine's'. Brain 1–10Hz, cerebellum 100Hz, joints 1kHz — every line of latency between the three layers must be owned by someone. This page gives a table of real, quotable numbers as of late 2026 plus five optimization levers — enough to quote numbers in an interview and match budgets on the line.

Latency number table: the 2026-11 reality

First, where the industry already is:
SystemHardwareLatency / frequencySource
π0 single chunkA single 4090About 80ms / chunkPhysical Intelligence public figures
π0 distilledSingle GPUUp to 50HzSame as above, after distillation
Xiaomi Robotics-0Consumer GPU80ms inference latencyOfficial release (2026)
StreamingVLA—271ms → 171msTsinghua public paper
SJTU 0.9B model—12.3Hz / 3.2GB VRAM / about 90% real-robot successPublic paper figures
How to read it: a single 80ms chunk ≈ a 12Hz brain — plenty for cerebellar interpolation; 'distilled to 50Hz' means the brain itself can connect directly to the mid-level controller. The point of this table is not 'fast enough' but that latency numbers have entered the usable range and the bottleneck has moved from compute back to data and generalization.

Frequency decoupling: three layers, three clocks

Deployed architectures are almost universally three-layer clocks: brain (VLA, 1–12Hz, emits action chunks) → cerebellum/policy layer (100–500Hz, interpolation within chunks, balance and compliance control) → joint/reflex layer (1kHz+, current loops, hard limits, e-stop — passing through no model at all). Helix's S2+S1 and GR00T's System 2/1 are the same principle productized: layers exchange only narrow-bandwidth signals (one intent vector or one chunk), each runs its own clock, none blocks the others. The line engineers most often get wrong: chunk length × execution frequency = the brain's fault-tolerance window. A 50-step chunk executed at 200Hz = 250ms between corrections — not enough budget for catching a fast-rolling object; either shorten the chunk or let the cerebellum layer correct online. This is not a hyperparameter; it's safety design set per task.

Five optimization levers (ordered by cost-effectiveness)

Compute tierWhat it can runWhat it cannot
Jetson Orin (on the robot)0.5–1B-class VLA + cerebellum control7B at full precision; large chunks at high frequency
Consumer GPU (4090-class)3–7B VLAs, high-frequency inference after distillationMulti-camera concurrent 7B
Cloud GPULarge-model brain layer (cloud architecture)Network loss and latency jitter are architectural risks (see <a href="/brain/">brain and cerebellum principles</a>)

Safety department: three things no model owns

E-stop loops, hard limits, and overcurrent protection are implemented at the MCU hardware layer — the model may err; the reflex may not. No amount of 'end-to-end' talk changes that this layer must exist independently.

Action-space filtering: chunks from the model pass through workspace/speed/force-constraint filters before being sent down — the model says 'move 80 centimeters within one second', the filter says no.

Model-timeout degradation: when the brain freezes, the cerebellum layer must fall back to safe behavior (stop / return to safe pose) without waiting. Embedded engineers are worth more, not less, in the large-model era — these three lines are why.

Sourcing: latency figures come from public releases by Physical Intelligence, Xiaomi, Tsinghua University, and SJTU; 'the bottleneck moved back to data' is a summary judgment as of 2026-11.

Previous · Post-trainingPost-training: from base model to your robot Next · Evaluation TruthEvaluation Truth: between 90% and 15%

All 6 teardown pages

01The model survey 02Taking a VLA Apart25 min03Data Engine: the real bottleneck of VLAs25 min04Post-training: from base model to your robot20 min05Deployment and Inference: landing a 1–10Hz brain20 min06Evaluation Truth: between 90% and 15%20 min