Cloud-Edge-Device: the compute revolution
Cloud thinks at 1-3 Hz; device acts at 1 kHz
The three-tier latency ledger
On-robot: under 10 ms for motion control, balance, collision avoidance - closed-loop control cannot wait for a network hop; hardware anchor NVIDIA Jetson Thor (2,070 FP4 TFLOPS at 40 W). Facility edge: 10-50 ms for fleet coordination, complex perception, VLA inference; JD's Asia No.1 edge cluster coordinates 500+ robots with sub-5 ms task assignment. Cloud: 50 ms+ acceptable for training, fleet learning, OTA; Tesla Dojo (D1 at 7 nm, 362 TFLOPS BF16, planned 100 EFlops).
| Tier | Latency target | What runs here | Typical hardware |
|---|---|---|---|
| On-robot | <10 ms | Motion control/balance/avoidance | Jetson Thor, embedded NPU |
| Facility edge | 10-50 ms | Fleet coordination/VLA inference | GPU edge servers |
| Cloud | 50 ms+ | Training/fleet learning/OTA | Dojo, public cloud |
The two-brain model: large model instructs, small model executes
Production VLAs in 2026 widely adopt two systems: System 2 (slow) an LLM at 1-2 Hz parsing language and emitting subgoals; System 1 (fast) a lightweight visuomotor policy at 30-50 Hz turning camera frames into motor commands. The key insight is System 2 need not run at control rate - hence aggressive quantization: LiteVLA with NF4 plus llama-cpp demonstrated working visuomotor control on a Raspberry Pi 4; the commercial standard is INT8/FP16 plus TensorRT on Jetson AGX Orin/Thor, with GR00T N1.7 shipping a dedicated Jetson deployment guide.
On-device small models (Phi-3.5-mini, Qwen2.5-0.5B class, under 1 B params, as low as 2 W) cover 80% of daily perception; the remaining 20% - non-standard grasps, exception cases - escalate to cloud GPT-4-class models. Unitree's consumer robots use exactly this split. Changjiang Securities' basis: on-device VLAs distilled below 10 B params achieve 1 kHz control while cloud 100 B-class models assist high-generalization scenes - long-term cloud-edge parallelism is the realistic path.
Data quality: the hidden multiplier of edge deployment
The counter-intuitive fact: the smaller the model, the more training-data quality matters. A 70 B cloud model brute-forces through noisy labels; a quantized 1 B edge model cannot - every mislabeled grasp point and every camera-action timestamp skew compounds into degradation at the edge. Measured basis: teams doing strict pre-training data preprocessing score 15-25% higher edge task success than teams training on raw captures. This is the real moat of AGIBOT/UBTECH heavy-asset data factories.
Deployment FAQ
If the factory loses network, can embodied robots still work?
Yes, degraded: on-device models keep executing issued subtask sequences and control; new language instructions and exception escalation wait for reconnect. Engineer task-package granularity for eight hours of offline continuation.