Multi-Lingual Voice Control System for AGV Management Using Edge LLMs

addverb · Tech · 2024 · source Agentic: yes · promotional bias: high · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Warehouse robotics maintenance

Addverb needed warehouse staff to resolve AGV issues through natural language across many languages without waiting for specialized maintenance engineers.

Constraints:

Warehouse workers may speak many different languages.
AGV downtime directly affects productivity.
Industrial environments need low latency and reliable local operation.
Wrong command translation can create operational or safety issues.

Architecture

Addverb built a multilingual voice-control assistant for Zippy AGVs using speech-to-text, Llama 3 on Supermicro/Intel edge servers for cost-effective local inference, and ChatGPT for cloud processing when higher performance is needed.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  speech-to-text-stage["Speech-to-Text Stage"]
  llm-router["LLM Router"]
  llama-3-edge-inference["Llama 3 Edge Inference"]
  chatgpt-cloud-processing["ChatGPT Cloud Processing"]
  agv-command-translator["AGV Command Translator"]
  supermicro-edge-server["Supermicro Edge Server"]
  zippy-agvs["Zippy AGVs"]
  speech-to-text-stage --> llm-router
  llm-router --> llama-3-edge-inference
  llm-router --> chatgpt-cloud-processing
  llm-router --> agv-command-translator
  llama-3-edge-inference --> llm-router
  llama-3-edge-inference --> supermicro-edge-server
  chatgpt-cloud-processing --> llm-router
  agv-command-translator --> llm-router
  agv-command-translator --> zippy-agvs
  supermicro-edge-server --> llama-3-edge-inference
  zippy-agvs --> agv-command-translator

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`Speech-to-Text Stage`	Converts spoken warehouse requests in supported languages into text.	`LLM Router`
`LLM Router`	Selects edge Llama 3 or cloud ChatGPT based on cost, latency, and performance needs.	`Llama 3 Edge Inference`, `ChatGPT Cloud Processing`, `AGV Command Translator`
`Llama 3 Edge Inference`	Handles cost-effective on-premises routine inference.	`LLM Router`, `Supermicro Edge Server`
`ChatGPT Cloud Processing`	Handles requests where higher performance or faster generation is needed.	`LLM Router`
`AGV Command Translator`	Maps interpreted requests into Zippy AGV commands.	`LLM Router`, `Zippy AGVs`
`Supermicro Edge Server`	Runs on-premises AI inference using Intel Xeon processors.	`Llama 3 Edge Inference`
`Zippy AGVs`	Warehouse sorting robots controlled and maintained through the assistant.	`AGV Command Translator`

Data flow

A worker speaks a maintenance request; speech-to-text converts it, the system routes the request to Llama 3 at the edge or ChatGPT in the cloud, and the interpreted instruction is translated into AGV commands.

“Addverb's solution employs a multi-stage pipeline that converts spoken language into AGV commands:”summary, stated

“The solution combines speech-to-text technology with a dual LLM architecture—using Meta's Llama 3 for cost-effective edge inference and OpenAI's ChatGPT for cloud-based processing”summary, stated

Why this architecture

The hybrid edge/cloud design balances local reliability, privacy, cost, and latency against the option to use a more capable cloud model for harder or faster-generation cases.

“This dual-LLM approach represents a practical architectural pattern where organizations balance cost, latency, and performance by routing requests to different models based on requirements.”summary, stated

“The choice of edge computing over pure cloud deployment reflects several considerations relevant to industrial LLMOps:”summary, stated

Models used: Meta Llama 3 · OpenAI ChatGPT

Scale

The assistant supports 98 languages.
“supporting 98 languages to accommodate the diverse workforce in modern warehouses.”summary, stated
Addverb serves over 350 clients.
“The company serves over 350 clients including major enterprises like Coca-Cola, Amazon, and DHL.”summary, stated
The selected edge server supports up to 32 cores and 2TB DDR5 RAM.
“Up to 32 cores (64 threads)”summary, stated

Alternatives considered (roads not taken)

Pure cloud LLM processing

Rejected because: Warehouse maintenance needs lower latency, local reliability, data locality, and lower routine query cost.

“Edge processing reduces dependency on consistent internet connectivity”summary, stated

Only specialized maintenance engineers

Rejected because: AGV fleets create maintenance demand and many issues could be resolved by untrained staff if communication were easier.

“Many AGV issues could potentially be resolved by untrained warehouse staff if they had a way to communicate with the machines effectively.”summary, stated

Patterns

`model-routing` (architecture)

Requests can use Llama 3 at the edge or ChatGPT in the cloud depending on operational needs.

“commercial cloud-based LLM is used when "performance is needed," offering "reduced time to generate the instructions (tokens)" sent to the AGVs.”summary, stated

`tool-use-orchestration` (architecture)

The language model output is converted into concrete AGV commands.

“translate unstructured verbal requests into AGV commands.”summary, stated

Tradeoffs

Hybrid edge/cloud LLMs

Chose Llama 3 locally plus ChatGPT when performance is needed over One model path for all requests.

Rationale: The system balances cost, latency, connectivity, and capability.

“balancing costs and capabilities.”summary, stated

Use speech as the control surface

Chose Multilingual natural speech over Training every worker on specialist maintenance tools.

Rationale: Natural speech lets untrained staff communicate with AGVs across languages.

“allows warehouse workers to communicate with machines in 98 different languages using natural speech.”summary, stated

Gotchas (what bit them)

Edge LLMs add distributed maintenance burden.

Root cause: Models and hardware must be updated and operated across warehouse edge sites.

Resolution: Use server hardware designed for AI/ML workloads and route harder cases to cloud processing.

“Deploying LLMs at the edge introduces operational challenges not present in cloud deployments, including hardware maintenance, model updates across distributed locations”summary, stated

Natural-language commands can be unsafe if translated incorrectly.

Root cause: Industrial systems need precise command generation from ambiguous speech.

Resolution: Keep command translation constrained and interpret vendor effectiveness claims cautiously.

“Incorrect translations could lead to operational issues or safety concerns in an industrial setting.”summary, stated

Evaluation

The case describes qualitative production goals but does not publish accuracy, latency, downtime, or cost metrics.

Metrics: language coverage · latency · downtime · cost

“No specific metrics are provided regarding accuracy rates for speech recognition or command translation, reduction in downtime, cost savings, or comparative performance”summary, stated

Outcomes

Addverb deployed an AI maintenance assistant for Zippy warehouse robots.
The assistant supports multilingual speech and edge/cloud model routing.
The published case does not provide hard performance metrics.

The case reports 98-language support, 350+ clients, and Supermicro edge servers with Intel Xeon processors.

“enabling on-premises AI inference to minimize downtime and increase productivity by allowing untrained warehouse staff to resolve AGV issues”summary, stated