How NetEase’s Ultra‑Low‑Latency Cloud Architecture Powers Remote Excavators

Recently, NetEase partnered with China Construction 8th Bureau on a strategic project in the high‑altitude railway construction zone of Litang, Sichuan, jointly developing excavation robots for plateau operations and applying gaming‑derived AI, interactive simulation, low‑latency audio‑video, and crowdsourced operations to the engineering machinery field.

The NetEase Fuxi excavation robot leverages NetEase Cloud Sign’s ultra‑low‑latency control signaling and media transmission, enabling remote phone control of excavators with end‑to‑end latency under 150 ms and multi‑view 1080p video streaming, even in poor outdoor network conditions.

Ultra‑Low‑Latency Global Transmission Architecture

Edge computing and edge access bring servers close to users. NetEase Cloud Sign has deployed edge nodes across major countries and many Chinese provinces, continuously monitoring network quality and using a “race” mechanism to keep the best servers near users.

Intelligent scheduling selects the optimal node based on five dimensions: static scheduling (nearest node), user login success rate, user business status (jitter, latency), real‑time probing (RTT, loss, jitter), and traffic aggregation (peak scheduling, load balancing).

To improve inter‑network transmission, NetEase Cloud Sign introduced the WE‑CAN (Communications Acceleration Network), a self‑developed large‑scale distributed transmission network that computes optimal public routes, avoids overlapping paths, applies QoS mechanisms, and quickly switches routes when network jitter or node failures occur.

WE‑CAN consists of four modules:

Scheduling node – allocates access nodes.

Access node – handles protocol conversion, service tiering, and hot updates.

Forwarding node – forms a full‑mesh network, reports RTT, loss, jitter.

Control node – collects reports and performs route planning.

Routing planning uses a link‑quality MOS score (0‑1) derived from RTT, loss, and jitter. The Dijkstra algorithm finds the shortest path, and multiple best paths are generated to avoid congestion. If a node exceeds traffic thresholds, traffic is shifted to secondary paths, and the process repeats to produce the final routing table.

Fast obstacle avoidance is achieved through three mechanisms: control nodes update routes upon large‑scale jitter or node failure, forwarding nodes employ ARQ and FEC to combat packet loss, and routes switch to secondary paths when RTT exceeds a threshold.

Ultra‑Low‑Latency Real‑Time Signaling

Traditional signaling and media channels are separated, leading to inconsistent weak‑network performance. NetEase’s solution multiplexes signaling and media over a single channel, using a self‑developed reliable UDP protocol on top of the WE‑CAN network. This provides average end‑to‑end latency of 75 ms, 100 % delivery reliability, and synchronized vehicle video, control commands, and status.

Ultra‑Low‑Latency QoS Techniques

Key QoS measures include limiting the jitter buffer’s target delay to 100‑200 ms, forcing frame output during the buffer’s convergence period, and optimizing rendering time by selecting the minimum of the original render time and a newly calculated render time (receive_frame_time + target_delay).

Both post‑decode and pre‑decode frame dropping are employed to eliminate excessively delayed frames, ensuring smooth playback without increasing latency.

Congestion control combines global smoothing (adjusting smoothing coefficients based on bandwidth availability) and frame‑level pacing (ensuring frame intervals do not exceed the interval plus 50 ms).

Transmission Quality Comparison

Comparisons between WE‑CAN and public internet links from China to the United States show higher quality transmission rates (over 95 % arrival) and lower RTT for WE‑CAN.