You can deploy this model and start earning money today!

lotapa/
mobile-actions-robot_q8_ekv1024.litertlm

Name: mobile-actions-robot_q8_ekv1024.litertlm
Availability: InStock

Views

No views yet

mobile-actions-robot_q8_ekv1024

A quantized on-device LLM fine-tuned for robot action command routing on Android devices.
It runs via Google LiteRT (.litertlm format) directly on mobile hardware — no internet or cloud inference required.

Model Summary

Property	Value
Format	LiteRT LLM (`.litertlm`)
Quantization	Q8
KV Cache	1024 tokens
Max tokens	512
Temperature	0.1
Top-K	32
Top-P	0.9
Backend	CPU / GPU (Android)

What This Model Does

This model is a robot action router. It takes a natural language instruction and calls the appropriate tool function to produce a structured robot command. It does not generate conversational text — its only job is to map user input to one of the supported robot actions via tool calling.

It is designed as the first stage in a two-model pipeline on Android:

This model — classifies the user's intent and dispatches a robot command via tool call
A chat model (e.g. Gemma) — handles everything else (general questions, conversation)

The smile action is the model's catch-all signal: when the input is not a recognized robot command, the model calls smile() and the host app falls through to the chat model for a conversational response.

Supported Actions

Tool	Action String	Description
`moveForward()`	`FORWARD`	Move the robot forward
`moveBackward()`	`BACKWARD`	Move the robot backward
`turnLeft()`	`LEFT`	Turn left in place
`turnRight()`	`RIGHT`	Turn right in place
`stop()`	`STOP`	Stop immediately
`turnAround()`	`TURN_AROUND`	Rotate 180°
`singASong()`	`SING`	Play a melody / sing
`smile()`	`SMILE`	Catch-all — not a robot command

System Prompt

You are a robot action controller. You MUST call a tool for every movement request —
never respond with text when movement is requested. Call moveForward for 'move forward',
moveBackward for 'go back', turnLeft for 'turn left', turnRight for 'turn right', stop
for 'stop'. Only skip tool calls for non-movement questions.

How to Use (Android / LiteRT)

1. Load the model

kotlin

val engine = Engine(
    EngineConfig(
        modelPath = "/path/to/mobile-actions-robot_q8_ekv1024.litertlm",
        backend = Backend.GPU,   // automatically falls back to CPU
        maxNumTokens = 512,
        cacheDir = context.cacheDir.absolutePath,
    )
)
engine.initialize()

2. Define your tools

kotlin

class RobotTools(val onCommand: (String) -> Unit) {

    @Tool(description = "Move the robot forward at the default speed for a short duration.")
    fun moveForward(): Map<String, String> { onCommand("FORWARD"); return mapOf("result" to "success") }

    @Tool(description = "Move the robot backward at the default speed for a short duration.")
    fun moveBackward(): Map<String, String> { onCommand("BACKWARD"); return mapOf("result" to "success") }

    @Tool(description = "Turn the robot left in place for a short duration.")
    fun turnLeft(): Map<String, String> { onCommand("LEFT"); return mapOf("result" to "success") }

    @Tool(description = "Turn the robot right in place for a short duration.")
    fun turnRight(): Map<String, String> { onCommand("RIGHT"); return mapOf("result" to "success") }

    @Tool(description = "Stop the robot immediately.")
    fun stop(): Map<String, String> { onCommand("STOP"); return mapOf("result" to "success") }

    @Tool(description = "Turn the robot 180 degrees to face the opposite direction.")
    fun turnAround(): Map<String, String> { onCommand("TURN_AROUND"); return mapOf("result" to "success") }

    @Tool(description = "Make the robot sing a song or play a melody.")
    fun singASong(): Map<String, String> { onCommand("SING"); return mapOf("result" to "success") }

    @Tool(description = "Make the robot smile or show a happy expression.")
    fun smile(): Map<String, String> { onCommand("SMILE"); return mapOf("result" to "success") }
}

3. Create a conversation and send a prompt

kotlin

val robotTools = RobotTools { action -> println("Command: $action") }

val conversation = engine.createConversation(
    ConversationConfig(
        samplerConfig = SamplerConfig(topK = 32, topP = 0.9f, temperature = 0.1f),
        systemInstruction = Contents.of(Content.Text("...system prompt above...")),
        tools = listOf(robotTools),
    )
)

// Send a user instruction
conversation
    .sendMessageAsync(Contents.of(Content.Text("move forward")))
    .collect { /* tool callback fires automatically */ }

Recommended Fallback

For robustness, complement the model with a simple keyword fallback for cases where it returns null:

Keywords	Action
`forward`, `ahead`, `straight`	`FORWARD`
`backward`, `back up`, `reverse`	`BACKWARD`
`turn left`, `go left`, `rotate left`	`LEFT`
`turn right`, `go right`, `rotate right`	`RIGHT`
`stop`, `halt`, `freeze`	`STOP`
`turn around`, `180`, `spin around`	`TURN_AROUND`
`sing`, `song`, `melody`, `tune`	`SING`

License

Gemma Terms of Use

lotapa/mobile-actions-robot_q8_ekv1024.litertlm

mobile-actions-robot_q8_ekv1024

Model Summary

What This Model Does

Supported Actions

System Prompt

How to Use (Android / LiteRT)

1. Load the model

2. Define your tools

3. Create a conversation and send a prompt

Recommended Fallback

License

lotapa/
mobile-actions-robot_q8_ekv1024.litertlm