Skip to main content

Cyberwave is in Private Beta.

Request early access to get access to the Cyberwave dashboard.

STUB DOCUMENT: This page is intentionally minimal and will be expanded with deeper technical details in a future update.

​
What “voice as a sensor” means in Cyberwave

Cyberwave treats every input stream as a sensor attached to a digital twin. A microphone is no different from a camera: audio is captured, optionally transcribed, and fed into the platform in one of three consumer roles.
Role of the transcriptWhat it drivesCyberwave surface it hits
InstructionA deployed ML / VLA controller policyPOST /api/v1/controller-policies/{uuid}/execute with {"instruction": "..."}
Trigger inputA workflowPOST /api/v1/workflows/{uuid}/execute
Telemetry / labelA recording, dataset, or alertcyberwave/twin/{uuid}/telemetry MQTT topic, twin.alerts.create(...)
Everything else (the wake word, the STT engine, the VAD) lives at the edge or on the client and is your choice.
Cyberwave never needs to hear audio directly. You decide where transcription happens; the platform only cares about the resulting text.

​
Architecture

Cyberwave exposes three stable primitives: controller-policy execute, workflow execute, and telemetry publish. Any voice stack that can produce text can drive a Cyberwave robot through them.

​
Two integration patterns

Best for operator-in-the-loop apps, demos, hackathons, and voice-first UIs.
Operator -> laptop mic -> faster-whisper (local) -> REST: execute controller policy
No changes on the edge device, no new sensor twin. Starts working in minutes.

​
Use cases

​
Manipulation (arm and gripper twins: SO-101, UR5, Piper, iiwa14, openarm, …)

Voice commandImplementation
Pick and place: “pick the red block and place it in the blue cup”instruction to a VLA or diffusion policy
Sort by attribute: “put all the pens in the left drawer”VLA multi-step, or a workflow with a per-object loop
Handover: “hand me the screwdriver”VLA plus force-sensitive gripper release primitive
Reset / home: “go home”, “reset workspace”Motion controller policy with a recorded sequence
Pause / resume / abort: “stop”, “continue”Workflow state control; controller-policy execute with a stop instruction, or MQTT cancel
Parametric tasks: “stack three blocks”, “pour 20 ml”Parse the number and pass it via workflow input rather than as a VLA prompt
Inspection: “show me the label of the bottle”VLA picks and rotates; camera twin captures the frame and posts to a workflow result
Labeling demonstrations: “this one is a successful pickup”Publish to telemetry topic; the episode trim tool picks it up
Teach-a-pose: “remember this as ‘tray‘“Records the current joint state under the given label for reuse in a motion controller

​
Locomotion (mobile twins: Go2, Husky-300, UR-mounted AMRs, humanoids)

Voice commandImplementation
Go-to: “go to the kitchen”, “return to dock”Navigation workflow with a named waypoint as input
Follow / come here: “follow me”, “come”Controller policy subscribing to person-detection sensor plus instruction
Speed and gait: “walk slower”, “trot”, “sit”Direct controller command (Go2 has built-in gaits)
Patrol / area scan: “patrol the hallway for 10 minutes”Workflow with a duration input; VLM-driven anomaly detection on frames
Fetch: “bring me that bottle”Locomotion policy plus handover, composed as a workflow
Emergency stop: “stop”, “freeze”Dedicated low-latency wake-word pipeline that publishes an abort command over MQTT
Exploration narration: “what’s in the corner of the room?”Locomotion to position plus VLM describe-scene plus TTS response

​
Cross-cutting (applies to both)

  • Recording control: “start recording”, “label this run ‘take 3’” drive cyberwave/twin/{uuid}/telemetry with telemetry_start events.
  • Mode switching: “take over” detaches the active controller policy and re-attaches Local Teleop.
  • Alert acknowledgement: “acknowledge all alerts” iterates twin.alerts.list(status="active") and calls alert.acknowledge() on each.
  • Status queries: “what’s the battery on the follower?” triggers a workflow that reads twin state and returns TTS.
  • Voice-labeled datasets: narration recorded while teleoperating becomes free-form natural-language labels on episodes, directly usable as VLA training prompts.

​
Where voice slots into the Cyberwave stack

The voice layer is intentionally outside the platform: you choose client-side vs on-robot, local vs cloud STT, English vs multilingual, wake-word style, and so on. Cyberwave’s side is three stable primitives: controller-policy execute, workflow execute, telemetry publish. Any voice stack that can produce text can drive a Cyberwave robot today.

​
Where to Go Next

Deploy a Controller Policy

Get a VLA or RL policy running on a twin so voice has something to drive.

Workflows

Route transcripts into multi-step orchestration with parameters and branching.

Perception Overview

See how audio fits alongside vision in the perception stack.