Edge Computing

Edge computing = doing the work on the device instead of in the cloud. For robots, this isn't an architectural preference — it's a hard requirement. A self-driving car can't wait 200ms to round-trip an image to AWS. By the time the response arrives, the obstacle is already a crash.

The latency budget

A typical perception → control loop in a mobile robot:

• Camera capture: 10–30ms
• Image processing + neural net inference: 30–100ms
• Path-planning / safety check: 10–50ms
• Motor command: 1ms

Total: ~50–180ms. Add a cellular network round-trip (50–200ms each way) and you're already in crash territory.

The hardware

• Raspberry Pi 4/5 — ₹5–8k. Good enough for ROS2 nodes and lightweight CV. No GPU.
• NVIDIA Jetson Nano — ₹15k. 4-core ARM + small GPU. Older, but still common.
• NVIDIA Jetson Orin Nano — ₹40-50k. 6-core ARM + 1024-CUDA-core GPU. 8W envelope. Runs YOLOv8 at 30+ FPS.
• NVIDIA Jetson AGX Orin — ₹1.5-2L. Production-grade autonomy compute.
• Google Coral — ₹6-10k. TPU-only, very low power, good for inference-only.
• Intel NUC + RealSense — Higher power, x86 compatibility. Used when you want to run desktop ROS2 stacks.

TensorRT — the secret sauce

NVIDIA's inference compiler. Takes a PyTorch / ONNX model and produces a device-specific optimized engine. Typical speedups:

• FP32 → FP16: 2× faster, near-zero accuracy loss
• FP16 → INT8 (with calibration): 1.5× faster, small accuracy hit

A YOLOv8n on Jetson Orin Nano: 8 FPS in PyTorch → 28 FPS as a TensorRT FP16 engine → 42 FPS at INT8. That's the difference between unusable and shippable.

In India

• CynLr (Bangalore) — vision-guided picking, runs everything on Orin AGX.
• Niqo Robotics (Bangalore) — Jetson Xavier on tractor-mounted sprayers.
• Ati Motors (Bangalore) — Jetson for warehouse-aisle perception.
• Detect Technologies (Chennai) — industrial visual inspection on the edge.
• Garuda Aerospace (Chennai) — Jetson on agricultural and inspection drones.

Indian companies prefer Jetson over Coral because NVIDIA's CUDA ecosystem means you can train and deploy from the same codebase.

See it in action

The Forge 03 lesson walks through running YOLOv8 on a Jetson in a ROS2 node — including the TensorRT conversion step.

Related concepts

[[neural-network]] · [[computer-vision]] · [[ros2]] · [[embodied-ai]]