The link will drop. The robot still needs to think.
A factory AGV behind a steel rack. An inspection drone inside a turbine hall. A pipeline crawler in a steel pipe. A maritime vessel below deck. The shared property: the radio link is intermittent or absent, and the robot still needs to perceive, decide, and act.
On-device perception
Camera frames, LiDAR scans, IMU streams flow through native inferenced. Object detection, anomaly detection, classification run on the robot's compute box — no round-trip to a base station.
Voice command
STT in, command parsed, action executed, TTS out — all on board. Useful for inspection crews talking to drones, factory operators directing AGVs, field technicians driving service robots.
xLSTM forecasting
The xLSTM forecasting workload handles trajectory forecasting, motion planning, and recovery-policy execution on time-series sensor data. Linear in sequence length — a better fit than transformers for hours-long telemetry windows.
Store-and-forward
The IoT worker buffers telemetry locally during disconnect windows and replays to fleet HQ when the link returns. No lost frames, no lost events.
Edge deployment
Runs on a Pi 4, an industrial gateway, a Jetson, or any Linux compute box on the robot. Same binary as the data-centre Eldric; only the activated modules differ.
Recovery decisions
When the link drops, a pre-loaded policy can drive the robot to a safe state. When the link returns, the fleet HQ reconciles. Soft real-time at the line; hard real-time stays the motor controller's job.