Does any voice data leave the device?

No. Recognition, the language model and synthesis all run in containers on the J40. There is no cloud account and no per-call cost.

How well does it follow commands, and what are the limits?

The 10 default gestures respond to flexible phrasing, because the language model matches intent. Each command is handled on its own with no memory of earlier turns. "Nod" may get only a spoken reply; "do a nod gesture" works around it.

Can one J40 control more than one arm?

One container drives one arm through one USB serial port. More arms need more containers and more ports.

How far can I stand, and what about noise?

The reSpeaker Flex is a four-microphone far-field array with the processed channel selected automatically. If the wake word triggers too easily or too rarely, adjust its threshold and cooldown in the device settings.

How do I add my own gestures?

Turn torque off, pose the arm by hand, capture frames, name the gesture and give it a trigger phrase, then save. Changes apply in about 30 seconds. The reply style is editable in the same settings page.

How do other systems connect to the arm?

Poll GET /observation on port 8765 of the J40 for the latest joint and gripper positions. A digital twin, recorder or safety monitor on the same LAN can read it directly.

What do I need to prepare before deploying?

A calibrated SO-ARM101 with motor IDs assigned, the reSpeaker Flex plugged in, a speaker, and internet on the J40 for the first boot. The deployment guide links the three one-time arm setup steps.

Voice-Controlled SO-ARM Robotic Arm

Talk to a SO-ARM101 robotic arm. Say the wake word and a command; the arm replies through a speaker and performs the matching gesture. Speech, the language model and arm control all run on one reComputer J40 (Orin NX 16GB), with no cloud API.

SenseCraftIntermediate30minAcoustic & Audio AI

reSpeaker Flex XVF3800 — the four-microphone array that hears the wake word and the command

Concept render of the SO-ARM voice controller

Reusable Site Templates

A desktop arm that people command by voice, with no teach pendant and no code.

Trade shows and showrooms — visitors say a command and watch the arm respond.
Robotics and AI teaching — one project covers voice, language models and an arm.
Voice layer for an arm project — other programs read the live joint state over the LAN.
Hands-free manipulation research — explore voice control for assistive use.

Not a fit: multi-step dialogue that depends on earlier commands, precise trajectories, or picking objects by sight. Each command is handled on its own, and this arm has no camera.

How it works

Voice-controlled SO-ARM device composition: reSpeaker Flex connects over USB to a reComputer J40, which drives the SO-ARM101 over USB, plays replies through a speaker and serves joint state to LAN applications over HTTP

Four things on one desk: a SO-ARM101 arm, a reSpeaker Flex microphone array, a speaker, and a reComputer J40 that runs everything. The arm and the microphone connect to the J40 over USB. Once the first start has downloaded images and models, nothing needs the internet.

reSpeaker Flex XVF3800 is a four-microphone far-field array that picks up the wake word and the command. Plug it in before deploying; the container selects it on start.
Any speaker plays the reply. The arm speaks before it moves, so the speaker is required.

The wake word is "Hey Jarvis". Its sensitivity and cooldown are adjustable in the device settings if it triggers too easily or too rarely.

How well it works

What the reComputer J40 + SO-ARM101 + reSpeaker Flex configuration does after deployment.

What you can use

Wake word, a spoken command, a spoken reply, then the gesture.
10 gestures ship by default: home, ready to pick, open and close gripper, look up and down, wave, nod, shake head and dance.
Phrasing is flexible: the language model matches intent, not exact words.
New gestures are taught by hand: turn torque off, pose the arm, capture frames, name it and give a trigger phrase. Saving takes about 30 seconds to apply.
The deployment page shows live joint values at 5 Hz.

Limits

Each command stands alone. Conversation history is cleared after every turn, so "do that again" has no context.
"Nod" may get a spoken reply without moving. Saying "do a nod gesture" works around it.
The wake word does not respond for about 30 seconds while saved changes apply.

What you get out of it

The arm container serves live state on port 8765 of the J40.

Interface	What it returns
`GET http://<jetson-ip>:8765/observation`	Latest position of the six joints and the gripper, as flat JSON
`GET http://<jetson-ip>:8765/observation/schema`	Field types and value ranges for the observation

Joint values are normalized after calibration: -100 to 100 for body joints, 0 to 100 for the gripper. Any program on the same LAN can poll them — a digital twin, a demo recorder, a safety monitor.

Porting It to Your Own System

The reusable unit is wake word → local speech recognition → local language model picks a gesture → arm executes → spoken reply. The gesture library and the prompt are data, not code.

Layer	What porting costs you
Microphone selection, wake word, speech recognition and synthesis	reuse as-is
Local language model service on the J40	reuse as-is
Gesture library	teach or edit in device settings
Reply style and trigger phrases	edit in device settings
Arm driver	rewrite for an arm other than SO-ARM101

Where this shape does not fit

Tasks that need vision — finding and picking a named object. See the voice-controlled grasping arm reference design, which adds a wrist depth camera.
Long multi-step instructions that rely on earlier turns.

How to deploy

Check the setup and its product list, then follow the deployment guide.

Deployment Options

Voice Control Your SO-ARM

Deploys voice-arm to Jetson to talk to your SO-ARM; poses/prompts are configurable.

3 steps · ~30min

Open the deployment guide →

Robotic Arm

so_arm

Microphone

respeaker_flex

Compute Unit

recomputer_j40

Voice Brain

recomputer_j40

FAQ

Yes, after the first start. The first boot downloads about 10 GB of images and models; after that speech, the language model and arm control run on the J40 with no online API.

Didn't find your answer?

We Are Glad to Be Your Hardware Partner !