Voice-Controlled Warehouse Management
Say it out loud and the stock movement is logged.
Voice-controlled warehouse management on edge hardware — speak the transaction, MCP writes it to inventory. Five tiers, from cloud trial to a fully offline Jetson deployment.
Reusable Site Templates
Warehouse work where hands are full, nothing can be typed, and 2.4 GHz Wi-Fi covers the floor.
Receiving and shipping docks
Third-party receiving docks, store back rooms, line-side stores. "Stock in 5 units" writes the receipt and reads back the total.
Forklifts and high racking
Aisles in high-bay storage, hardware racking. The driver stays seated and asks what is on the shelf.
Line-side stores and tool cribs
Assembly line-side material areas, mould stores, tool cribs. Issues and returns are called out as they happen, with a daily summary at shift change.
Not a fit: dictating long batch numbers or barcodes, operations needing several turns of dialogue, and constant high noise or respirator use. The interface has no offline queue, so writes issued while the network is down are not replayed.
How it works
A SenseCAP Watcher captures the speech and the speech AI turns it into a command, either in the SenseCraft cloud or on a local reComputer J40 Series or reComputer Robotics J50 Series. A reComputer edge gateway writes the record and exposes it to the ERP or WMS you already run.
Four device domains, read left to right: voice terminal, speech AI, edge gateway, existing system. Only the speech AI has a placement choice — cloud, or one of two on-premise Jetsons — and that single choice is what produces the five deployment tiers below.
Voice Front-End
The Watcher carries two processors and two separate firmware images, flashed over USB-C:
| Processor | Firmware | Role |
|---|---|---|
| ESP32 | Xiaozhi voice agent | Wake word, audio capture, audio streaming, MCP client |
| Himax WE2 | Face recognition image | On-device face detection for operator verification |
Wi-Fi is 2.4 GHz only — a 5 GHz-only SSID will not appear in the setup scan. On-device face templates are capped at 20; beyond that, recognition moves to the gateway.
Edge Gateway
A Docker service on the reComputer gateway, source at github.com/suharvest/warehouse_system. It holds inventory state, serves a browser dashboard, and publishes an MCP endpoint.
The MCP endpoint is the integration contract. Everything the voice side can do is a tool on that endpoint, which means:
- an existing ERP/WMS is bridged by mapping its data interface onto the same tool surface — on the order of a few dozen lines of mapping code, not a migration
- sites without a WMS use the built-in one, including bulk inventory import from an Excel template
- the command vocabulary is extended by adding tools, with no change to the Watcher firmware
Speech AI
Speech recognition, the LLM and speech synthesis run in one of three places, and that placement is what separates the tiers:
| Placement | Device on site | LLM | Speech | Network | Tiers |
|---|---|---|---|---|---|
| SenseCraft cloud | None | Seeed-hosted | Seeed-hosted | Internet required | Trial, Tier 1, Tier 2A |
| Local speech, LLM by API | reComputer J4012, serving 1–3 Watchers | Your OpenAI-compatible endpoint (DeepSeek, etc.) | On the J4012 | Internet required for the LLM | Tier 2B |
| Fully on premise | reComputer Robotics J5011 on the LAN | EdgeLLM — Qwen3.5-4B | OpenVoiceStream on the J5011 (recognition, synthesis, voiceprint) | LAN only | Tier 3 |
On Tier 3 the LLM and the speech service both run on the J5011, and nothing calls an outside service after deployment. The J5011 needs about 10 GB of free disk for the speech models and the LLM engine. Measured local inference on the reComputer Robotics J5011: ~16 tokens/sec. Container images, ports and engine profiles are listed in the engineering wiki.
The data interface is listed under What You Get below.
How well it works
What the warehouse team uses in the browser dashboard and on a SenseCAP Watcher once the gateway is running.
Measured results
Benchmark and capacity figures measured on reference hosts across warehouse service and edge AI workloads.
| Metric | Typical | Device | Deployment Note |
|---|---|---|---|
| On-premise LLM inference | Qwen3.5-4B 8k engine (fully offline) | reComputer J40 Series (J4012) | JetPack 6.2 / TensorRT 10.3 acceleration, local voice closed loop |
| Material list query latency | P50 14.8 ms (P99 22.1 ms) | reComputer R2000 Series | 10 concurrent requests, SQLite atomic transactions over Tailscale |
| Stock transaction rate limit | 600 req/min (anti-abuse throttle) | reComputer R2000 Series | Counted per authenticated identity to safeguard core inventory jobs |
Known limit: the interface has no offline queue. Writes issued while the network is down fail and are not replayed automatically. For full API stress benchmarks and local LLM deployment, see Engineering Wiki · Voice Warehouse Management.
What you get out of it
Once the gateway is up, these interfaces are reachable from your own LAN at the reComputer's address.
| Interface | Endpoint / port | What you get | Which tiers |
|---|---|---|---|
| HTTP dashboard + REST API | http://GATEWAY_IP:2125 | Inventory list, transaction history, agent connection status | Every tier with a gateway (Tier 1 and above) |
| HTTP health check | http://GATEWAY_IP:2125/health | 200 once the warehouse service is ready; wire it into your monitoring | Tier 1 and above |
| MCP endpoint | port 18004 | The integration contract: every inventory operation is an MCP tool | Self-hosted on Tiers 2B and 3; issued by SenseCraft on Trial / Tier 1 / Tier 2A |
| Inventory database | SQLite /data/warehouse.db | Mount or dump it to feed an external BI or ERP job | Tier 1 and above |
| Voice server | WebSocket 18000, admin console 18003 | Watcher audio stream and its administration | Tiers 2B and 3 |
| On-premise AI | 8000 (EdgeLLM), 8621 (OpenVoiceStream) | Qwen3.5-4B behind an OpenAI-compatible API; recognition, synthesis and voiceprint | Tier 3 |
The MCP endpoint is where an existing ERP or WMS connects. Every voice action is a tool on it, so adding an action means adding a tool.
Porting It to Your Own System
The reusable unit is not "warehousing" — it is the chain voice entry point → MCP tools → business system. Only the last link is domain-bound: the warehouse system registers stock-in, stock-out and inventory lookup as MCP tools. Swap those for actions your own system already exposes and the first two links stay as they are.
| Layer | What porting costs you |
|---|---|
| SenseCAP Watcher and its firmware | reuse as-is |
| Edge gateway, deployment tiers, AI layer placement | reuse as-is |
| MCP endpoint and how the agent attaches | reuse as-is |
| The MCP tools your business system registers | build against your own interface — comparable to bridging one ERP/WMS |
| Agent role template — prompt, response language, command aliases | configuration |
| Data model and bulk-import template | configuration |
Fits: production reporting, equipment inspection, material calls, work-order sign-off, patrol logging, laboratory sample intake — tasks where hands are occupied, one operation carries three or four fields at most, and an existing system accepts writes through an interface.
Where this shape does not fit
- Long exact codes spoken aloud — batch numbers, serial numbers, barcodes. One misrecognised character costs more rework than the time saved.
- A single operation spanning several turns — complex context and branching held across a dialogue.
- Sustained noise or respiratory protection — pickup distance and wake-word reliability drop.
How to deploy
Tell us about your site — we work out the hardware, then you connect the data and follow the steps.
Configure yours
Answer three questions about the site — the hardware follows.
Operator Identity
Do transactions need to record who did them?
Configure yours
Answer three questions about the site — the hardware follows.
Trial · Starter
Watcher only, no host to install. The warehouse database and the voice AI are hosted by Seeed and billed monthly. No ERP/WMS integration and no face recognition.
硬件/物料名称 数量 型号/SKU 说明 SenseCAP Watcher 1 113991275 - Tier 1 · Basic
Inventory data stays on your own gateway; the voice path runs against SenseCraft cloud, whose cost is already included with the Watcher. Operator identification is on-device, up to 20 face templates.
硬件/物料名称 数量 型号/SKU 说明 SenseCAP Watcher 1 113991275 - reComputer R1125-10 1 113991334 - Tier 2A · Advanced (Single Site)
Industrial gateway with a Hailo-8 accelerator: warehouse system plus gateway-side face recognition with liveness detection. The speech path still calls SenseCraft or an OpenAI-compatible endpoint you supply.
硬件/物料名称 数量 型号/SKU 说明 SenseCAP Watcher 1 113991275 - reComputer Industrial R2135-12 (Hailo-8) 1 100061305 - Tier 2B · Advanced (Multi Site)
One Jetson carries the whole site — warehouse system, face recognition and the self-hosted speech service — shared by up to 3 Watchers. Only LLM calls go to the provider you choose.
硬件/物料名称 数量 型号/SKU 说明 SenseCAP Watcher 3 113991275 - reComputer Super J4012 1 114110314 - Tier 3 · Premium
Pure LAN. The industrial gateway runs the warehouse system and face recognition; the Jetson runs EdgeLLM (Qwen3.5-4B) and OpenVoiceStream, so nothing leaves the facility and there is no recurring service cost.
硬件/物料名称 数量 型号/SKU 说明 SenseCAP Watcher 1 113991275 - reComputer Industrial R2135-12 (Hailo-8) 1 100061305 - reComputer Robotics J5011 1 100076722 -
Or compare them yourself: preset specificationsDeployment Guide · 30min →
Hardware Equipment List by Preset
| Setup | Role | Device | Qty |
|---|---|---|---|
| Trial · Starter | Voice Assistant | SenseCAP Watcher | 1 |
| Tier 1 · Basic | Voice Assistant | SenseCAP Watcher / reTerminal D1001 8" HMI Display(either one) | 1 |
| Edge Gateway | reComputer R1100 Series | 1 | |
| Tier 2A · Advanced (Single Site) | Voice Assistant | SenseCAP Watcher / reTerminal D1001 8" HMI Display(either one) | 1 |
| Edge Gateway | reComputer Industrial R21 Series / reComputer J40 Series (Jetson Orin NX)(either one) | 1 | |
| Tier 2B · Advanced (Multi Site) | Voice Assistant | SenseCAP Watcher | 3 |
| Edge Server | reComputer J40 Series (Jetson Orin NX) | 1 | |
| Tier 3 · Premium | Voice Assistant | SenseCAP Watcher / reTerminal D1001 8" HMI Display(either one) | 1 |
| Edge Gateway | reComputer Industrial R21 Series | 1 | |
| AI Compute | reComputer J50 Series (Jetson AGX Orin) | 1 |
Technical Specifications & Implementation Comparison
| Tier | Devices | Network | AI location | Face recognition | Recurring cost |
|---|---|---|---|---|---|
| Trial · Starter | Watcher | Internet | Seeed cloud | ✗ | Monthly subscription |
| Tier 1 · Basic ⭐ | Watcher + reComputer R1125-10 | Internet | SenseCraft cloud | On-device only (≤20) | None — voice included with Watcher |
| Tier 2A · Advanced (Single Site) | Watcher + reComputer R2135-12 | Internet | SenseCraft cloud | Gateway-side, with liveness | No API key — SenseCraft account |
| Tier 2B · Advanced (Multi Site) | Watcher ×1–3 + reComputer J4012 | Internet | Local speech on the J4012; LLM via your API | On the J4012, with liveness | Your LLM provider |
| Tier 3 · Premium | Watcher + reComputer R2135-12 + reComputer J5011 | LAN only | On-premise Jetson | Gateway-side | None on the fully local LLM path; the private-cloud LLM option uses your LLM key |
Tier 2A supports two gateway-side inference backends for face recognition: Hailo-8 or Jetson.
The voice terminal in Tier 1, Tier 2A and Tier 3 is either a SenseCAP Watcher (default) or a reTerminal D1001, an 8" touch HMI with a camera. Both run the same voice firmware and reach the warehouse system over an MCP endpoint.
Difficulty is rated beginner; a Tier 1 deployment takes about 30 minutes.
FAQ
Tier 3 does. It runs the WMS, EdgeLLM (Qwen3.5-4B) and OpenVoiceStream on a Jetson inside the facility, so no inventory data or audio leaves the LAN. All other tiers depend on a cloud voice path.
Yes, from Tier 1 upward. The warehouse system exposes MCP tools; bridging means mapping your system's data interface onto that tool surface — roughly a few dozen lines of code. Tier 0 does not support integration.
No menu training. Operators speak the transaction; fuzzy matching resolves near-miss phrasings to the closest registered command. Deployment and gateway configuration do require a technician.
Cloud tiers are configurable — SenseCraft's hosted model, or any OpenAI-compatible endpoint you supply. The on-premise tier runs a fixed Qwen3.5-4B on EdgeLLM.
Face recognition. Under 20 operators, the Watcher's Himax WE2 handles it on-device. Above that, recognition runs on the gateway with a Hailo-8 or Jetson backend, which also adds liveness detection.
The voice command vocabulary (add MCP tools), the agent's name and response language, the LLM and TTS providers, and the inventory data model. Hardware customisation — enclosure, branding, I/O — is handled separately; see /customization.
Didn't find your answer?
Contact Us
We Are Glad to Be Your Hardware Partner !