Do voice or camera images leave the device?

No. Recognition, the language model, synthesis and detection all run on the J40. The camera frame is only served on the LAN through the dashboard.

How fast is it, and what can it pick up?

A full cycle from the voice command to carrying the object back takes about 11 seconds on reComputer J40. It grasps boxes, opaque cups, standing opaque bottles, bananas and oranges. Transparent bottles and objects wider than 0.100 m are out of range.

How should objects be placed?

Put one object 25–30 cm in front of the arm, straight ahead or moderately to one side. If it says it cannot find the object, move it toward the center and repeat the command.

Can it recognize objects that are not in the default list?

Yes, for similar shapes. The detector takes its object vocabulary as a runtime input, so adding an object is a config edit with no retraining. Grasping still follows one of four shape routes — box, cylinder, long or round.

Why is hand-eye calibration needed, and how long does it take?

How the camera sits on the wrist differs per unit, and grasping needs that geometry to the millimeter. Calibration is done once and takes about 30 minutes with a printed marker board. A 1 mm error in the measured marker size shifts grasps by about 1 cm.

How do other systems connect to it?

Poll port 8775 for the arm position and gripper state, or port 8776 for the full state and the latest camera frame. Any program on the same LAN can read them.

Can I change the wake word, voice or language model?

The wake word is "Hey Jarvis" by default. The language model is a local Qwen3.5-4B served on port 8000. Replacing either means editing the compose configuration on the J40.

Voice-Controlled Grasping Arm (reBot B601-DM)

Say "Hey Jarvis, grab the water bottle" and a reBot B601-DM arm finds the object with its wrist depth camera, picks it up and brings it back in about 11 seconds. Speech, the language model, detection and grasp planning all run on one reComputer J40 (Orin NX 16GB), with no cloud API.

SenseCraftAdvanced60minAcoustic & Audio AIAI Agent Ready

Reusable Site Templates

A desktop arm that picks up a named object when someone asks for it.

Showrooms and exhibitions — visitors name an object on the table and the arm fetches it.
Robotics research and teaching — voice, vision, grasp planning and arm control in one working stack.
Fetch-and-hand-over prototypes — a starting point for assistive or service arms.
Adding new objects — the list of recognizable objects is a config edit, with no retraining.

Not a fit: transparent objects, objects wider than the 0.100 m gripper, cluttered bins, and production pick-and-place with a stated success rate.

How it works

Four things: a reBot B601-DM arm, an Orbbec Gemini 2 depth camera on its wrist, a reSpeaker USB microphone with a speaker, and a reComputer J40 that runs everything. All of them connect to the J40 over USB. After the first start downloads the models, nothing needs the internet.

reBot B601-DM — a 6-joint arm with a force-controlled parallel gripper that opens to 0.100 m. It connects over USB serial.
Orbbec Gemini 2 — a stereo depth camera mounted on the wrist. It needs a USB 3.0 port; on USB 2 the depth stream stalls.

Hand-eye calibration is required once per unit before grasping. It takes about 30 minutes: print a marker board, let the arm sweep about 16 poses over it, and save the result. Until then, voice, detection and the dashboard work, but grasp commands do not move the arm.

reComputer J40 series (Orin NX 16GB) runs three services from one compose file:

Service	What it does
Voice	Streaming speech recognition and speech synthesis
Language model	A local Qwen3.5-4B that turns the sentence into a grasp request
Arm agent	Wake word, camera, object detection, grasp planning, arm control and the live dashboard

Plan for about 10 GB of free disk. The first start downloads about 8.5 GB of models on top of 1.4 GB of container images.

The data interface is listed under What You Get below.

How well it works

Typical values from the deployment guide, on reComputer J40 + reBot B601-DM + wrist Gemini 2, after hand-eye calibration.

What the user gets	Typical	Device
Voice command to object carried back	about 11 s	reComputer J40 series (Orin NX 16GB)
Object types it grasps	5 — boxes, opaque cups, standing opaque bottles, bananas, oranges	reComputer J40 series (Orin NX 16GB)

Limits

Transparent bottles cannot be grasped. The stereo depth camera does not see clear plastic or water.
Objects wider than 0.100 m are declined with a spoken reply.
Best placement is 25–30 cm in front of the arm, straight ahead or moderately to one side.
If the detector engine does not load, the stack falls back to a slower build of the same model; detections stay the same.

What you get out of it

All interfaces are served by the J40.

Interface	What it carries
`http://<jetson-ip>:8776`	Live dashboard: wrist-camera view and arm state
`GET http://<jetson-ip>:8776/api/state`	Arm pose, gripper state and the last grasp result, as JSON
`GET http://<jetson-ip>:8776/api/frame.jpg`	Latest wrist-camera color frame
`GET http://<jetson-ip>:8775/observation`	Arm position x / y / z and whether the gripper is holding something
`http://<jetson-ip>:8000`	The local language model service

Other systems read state; they do not need to drive the arm. A recorder, digital twin or line monitor can poll the observation endpoint on the LAN.

Porting It to Your Own System

The reusable unit is wake word → local recognition → language model picks the object → detect it in the wrist camera → plan and execute the grasp → spoken result.

Layer	What porting costs you
Voice, wake word and the local language model	reuse as-is
Object list for the detector	edit the config — no retraining
Grasp routes for boxes, cylinders, long and round objects	reuse for similar shapes
Hand-eye calibration	redo for every unit
Arm driver	rewrite for an arm other than reBot B601-DM

Where this shape does not fit

Transparent or reflective objects — the depth camera cannot see them.
Bin picking and dense clutter — the design expects separated objects on a table.
Throughput or success-rate commitments — measure on your own objects and layout first.

How to deploy

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

Hardware Sizing & Deployment Setup

Select site conditions to generate recommended bundles and guides, or let an AI assistant estimate hardware and deploy for you.

Self-service Consult & Deploy

Deployment Options

Voice Grasping on Jetson

One compose file deploys voice+LLM+arm to Jetson Orin NX; calibration unlocks grasping.

3 steps · ~60min

Deployment Guide →

Robotic Arm

rebot_arm_b601

Depth Camera

Orbbec Gemini 2 RGB-D Camera

Microphone

respeaker_flex

Compute Unit

recomputer_j40

FAQ

Yes, after the first start. The first boot downloads about 8.5 GB of models and 1.4 GB of images; after that speech, the language model, detection and grasping run on the J40 with no online API.

Didn't find your answer?

We Are Glad to Be Your Hardware Partner !