functiongemma-270m-drone-router-navlink-v2

FunctionGemma 270M Drone Router — NAVLINK Round 2 (Reproduced Checkpoint)

This repository contains the reproduced round-2 LoRA adapter checkpoint for NAVLINK in:

• Model repo: llama-farm/functiongemma-270m-drone-router-navlink-v2
• Dataset snapshot: llama-farm/drone-router-dataset-navlink-v2
• Verified reproduced eval: 95.0% exact tool-call accuracy (475/500) with 0 safety violations
• Original canonical best (overwritten): 95.2% (476/500), 0 safety violations

This upload is the reproduced round-2 checkpoint rebuilt from the restored 4,307-example dataset after the original adapter weights were overwritten.

What the model expects

The model is a compact command router for PX4/MAVLink drone control. It expects a normalized two-line textual interface:

[STATE] <telemetry + vision context>
[CMD] <operator / vision / safety trigger>

It returns:

drone_tool(param="value")
Short operator-facing feedback.

Heartbeat exception

For heartbeat-only notifications, the model may emit a single line only:

drone_notify(message="heartbeat")

Exact system prompt

You are a drone command router. Given [STATE] telemetry and a [CMD] trigger, output exactly one tool call on line 1, and a short operator confirmation on line 2.

Safety rules (override everything):
- battery < 15%: drone_flight(action="land")
- battery < 25%: drone_flight(action="rth")
- gps lost: drone_flight(action="stop")

Routing hints:
- Target not visible + "find X": drone_scan(lock_on_class="X")
- Target visible + "follow/track": drone_track(action="start")
- Multiple waypoints: drone_mission, single GPS point: drone_goto
- Compound commands: first step only

Heartbeats: drone_notify(message="heartbeat") — no line 2.

Exact [STATE] + [CMD] interface

Base format

[STATE] phase=AIRBORNE alt=15.2m bat=72% hdg=270° spd=3.1m/s pos=32.851,-97.125 | vision=SCANNING target=person wp=2/5
[CMD] <trigger text>

With active detection context

[STATE] phase=AIRBORNE alt=15.0m bat=68% hdg=270° spd=0.0m/s pos=32.852,-97.124 | vision=SCANNING target=person wp=0/0 | det=person conf=0.87 bearing=45° range=15m at=32.853,-97.121
[CMD] person detected at bearing 45°, confidence 87%

Telemetry field format

Core telemetry fields

• phase — uppercase flight phase such as GROUNDED, AIRBORNE, RTL
• alt — altitude above ground, formatted like 15.2m
• bat — battery percentage, formatted like 72%
• hdg — heading in degrees, formatted like 270°
• spd — horizontal speed, formatted like 3.1m/s
• pos — GPS position as lat,lon

Vision fields

• vision — one of IDLE, SCANNING, TRACKING, LOST
• target — none, person, vehicle, animal, or any when appropriate
• wp — waypoint progress as N/M (or 0/0 if not in a waypoint mission)

Optional detection fields

• det — detected class
• conf — confidence score, typically 2 decimals (example: 0.87)
• bearing — relative bearing in degrees (example: 45°)
• range — estimated range in meters (example: 15m)
• at — estimated target GPS as lat,lon

Vision trigger formatting

Vision events should be normalized into the exact same [STATE] + [CMD] interface.

Examples:

[CMD] person detected at bearing 45°, confidence 87%
[CMD] possible vehicle, low confidence 38%
[CMD] lost tracking on person
[CMD] search complete, 2 targets found
[CMD] search complete, no targets

Map / UI / chat normalization guidance

All upstream inputs should be reduced into the same plain-text trigger format before inference:

• Chat / radio / typed pilot commands → keep the command terse and literal in [CMD]
• Ground-station map clicks → convert coordinates directly into the command text, e.g. [CMD] go to 32.853, -97.121 and scan
• UI button actions → normalize to the same canonical wording used in training where possible
• Vision triggers → include fused detection context in [STATE] and the event sentence in [CMD]
• Safety triggers → emit explicit command text such as [CMD] battery low at 22% or [CMD] gps signal lost

The model performs best when all surfaces—map, UI, chat, and vision—are normalized into the same textual contract instead of using separate APIs.

Output contract

Standard case

Line 1 must be exactly one executable tool call.

Line 2 should be a short operator-facing confirmation.

drone_tool(param="value")
Short operator-facing confirmation.

Heartbeat one-line exception

drone_notify(message="heartbeat")

Evaluation contract

• Line 1 is the hard interface and is scored by exact string match.
• Line 2 is helpful but non-executable operator feedback.

Safety rules

The model is trained with strict safety overrides:

• battery < 15% → drone_flight(action="land")
• battery < 25% → drone_flight(action="rth")
• gps lost → drone_flight(action="stop")

These rules take precedence over normal mission continuation, tracking, or convenience behavior.

Tool surface

The router targets 9 tools:

• drone_flight(action)
• drone_move(direction, distance_m)
• drone_yaw(degrees|heading)
• drone_goto(lat, lon, alt_m, speed_mps, on_arrival)
• drone_look(action)
• drone_scan(lock_on_class)
• drone_track(action, target_class)
• drone_mission(action, target, pattern, waypoints)
• drone_notify(message)

Known limitation

The main remaining weakness in this reproduced checkpoint is exact waypoint copying for multi-waypoint mission prompts. The model can still normalize or hallucinate waypoint lists when asked to patrol/search explicit coordinates in free text.

Operational guidance:

• prefer structured UI/map coordinate injection when possible
• validate generated waypoint strings before execution
• treat multi-waypoint free-text prompts as lower-confidence than single-point drone_goto(...) routing

Files in this repo

• adapter_config.json
• adapters.safetensors
• intermediate checkpoint adapters from the reproduced MLX run
• eval_results.json with the verified reproduced scorecard

Verified reproduced result

accuracy = 0.950
correct = 475
total = 500
safety_violations = 0

Note: The reproduced run scored 475/500 (95.0%) vs. the original 476/500 (95.2%). The 1-sample difference is within normal training variance. The same failure patterns persist: waypoint exact-copying in multi-point missions and occasional repetition in notify messages. See eval_results.json for the full per-tool breakdown and failure details.