Cloud-edge-device: the compute revolution of embodied AI
Cloud thinks at 1–3 Hz, device acts at 1 kHz
Three-layer latency ledger
On-device (on-robot): <10 ms, running motion control, balance and obstacle avoidance — closed-loop motion control cannot afford a network round trip; hardware anchor: NVIDIA Jetson Thor (2070 FP4 TFLOPS, 40 W). Factory edge (on-site): 10–50 ms, running fleet coordination, complex perception and VLA inference; JD's Asia No.1 edge cluster coordinates 500+ robots with task-assignment latency <5 ms. Cloud: 50 ms+ acceptable, running training, fleet learning and OTA; Tesla Dojo (D1 7 nm, BF16 362 TFLOPS, planned 100 EFlops).
| layer | Latency target | What it runs | Typical hardware |
|---|---|---|---|
| On-device | <10ms | Motion control / balance / obstacle avoidance | Jetson Thor, embedded NPU |
| Factory edge | 10-50ms | Fleet coordination / VLA inference | GPU edge server |
| Cloud | 50ms+ | Training / fleet learning / OTA | Dojo, public cloud |
Dual-brain model: the large model issues commands, the small model executes
Production-grade VLAs in 2026 commonly use a dual-system design: System 2 (slow thinking), an LLM at 1–2 Hz, interprets natural language and decomposes subgoals; System 1 (fast acting), a lightweight visuomotor policy at 30–50 Hz, translates camera frames into motor commands. The key insight is that "System 2 does not need to run at control frequency" — enabling aggressive quantization: LiteVLA runs visuomotor control on a Raspberry Pi 4 with NF4 quantization + llama-cpp; the commercial standard is INT8/FP16 + TensorRT on Jetson AGX Orin/Thor, and GR00T N1.7 ships with a dedicated Jetson deployment guide.
On-device small models (Phi-3.5-mini, Qwen2.5-0.5B class, under 1 billion parameters, power as low as 2 W) handle 80% of daily perception; the remaining 20% — non-standard grasps and abnormal conditions — escalate to a cloud GPT-4-class model. Unitree's consumer robots use exactly this scheme: on-device handles navigation/voice/simple recognition, cloud handles complex command understanding and multi-turn dialogue. Per Changjiang Securities: on-device VLA distilled below ten billion achieves 1 kHz motion control, while a cloud hundred-billion model assists high-generalization scenarios — long-term cloud-edge parallelism is the realistic path.
Data quality: the hidden multiplier of edge deployment
A counterintuitive fact: the smaller the model, the more training-data quality matters. A 70-billion cloud model can brute-force its way past noisy labels; a quantized 1-billion edge model cannot — every mislabeled grasp point and every camera-action timestamp misalignment compounds at the edge into performance degradation. Measured: teams that do strict data preprocessing before training achieve 15–25% higher edge-task success than teams using raw collected data. This is the true moat of AgiBot/UBTECH's asset-heavy data factories.
Deployment Q&A
If the factory network goes down, can embodied robots still work?
Yes, but degraded: the on-device small model keeps executing the already-issued subtask sequence and motion control; new natural-language commands and exception escalation must wait for the network to return. Engineering-wise, task-package granularity should be designed for "8 hours of offline continuation."