Skip to main content

​
What is Perception?

Perception is how your robot understands its environment: what’s in the camera frame, where obstacles are, what objects to interact with, and what a human operator is asking for. On Cyberwave, perception data flows from sensors on the edge through digital twins into AI models, and back into your control loop in real time.
Perception models on Cyberwave can run in the cloud (good for heavy VLMs and batch inference) or on the edge through Edge Workers (good for low-latency, video-never-leaves-the-device use cases).

​
Hardware that Supports it

Cameras

USB webcams, laptop cameras, and IP cameras for live streaming, vision workflows, and dataset recording.

LiDAR & Depth Sensors

Connect 3D sensors via ROS, drivers, or custom integrations.

Onboard Robot Sensors

Wrist-mounted cameras, depth sensors, and IMUs that stream alongside the robot’s joint state.

Edge Compute

Run inference on Raspberry Pi, Jetson, or any Linux box co-located with your hardware.

​
How You Build It

​
1. Stream sensor data into a twin

Cameras and sensors paired through the Edge Core appear as live streams on their digital twin. View them in the dashboard, record them as datasets, or pipe them into a model. The wire format is consistent across cloud and edge.

​
2. Pick the right model

InputModel typeReference
Single imageVLM (vision-language)ML Models
Video streamVLM with temporal contextML Models
Image + stateVLA (vision-language-action)ML Models
CustomAny registered modelML Models
from cyberwave import Cyberwave

cw = Cyberwave()
camera = cw.twin("wrist-camera-01")
model = cw.model("vlm/qwen-vl")

frame = camera.latest_frame()
answer = model.ask(frame, prompt="Where is the red block?")
print(answer)

​
3. Run inference where it makes sense

For low-latency loops (closed-loop visual servoing, safety-critical detection), deploy your model as an Edge Worker so frames never leave the device. For heavier reasoning (planning, multi-modal VLMs), call the cloud-hosted model directly.

​
4. Wire perception into workflows

The camera_frame trigger in Workflows runs your model on every new frame and routes the output into downstream nodes, perfect for event-driven detection, anomaly alerts, or visual triggers for manipulation tasks.

​
Where to Go Next

Camera Quickstart

Stream your first camera into a digital twin.

Edge to Cloud VLM

Tutorial: run a VLM on edge frames and trigger cloud workflows.

Edge Workers

Deploy models alongside your hardware for low-latency inference.