sollaholla/robocor-local-agent

Robocor local agent (450M, v1)

The offline Agent model that runs on Robocor robots. It is LiquidAI/LFM2.5-VL-450M fine-tuned to turn short operator commands into robot tool calls, and it describes pictures from the robot's camera with the stock vision projector.

The Robocor app downloads these files itself the first time the robot needs them. Nothing here has to be installed by hand.

Files

File	What it is	Size
robocor-agent-450m-v1-Q4_0.gguf	Fine-tuned language model, Q4_0	219 MB
mmproj-robocor-agent-450m-v1-Q8_0.gguf	Vision projector, Q8_0 (unchanged from the base model)	103 MB
llama-server-v1-linux-arm64-cuda.tar.gz	llama-server for NVIDIA Jetson (JetPack 6, CUDA 12.6)	86 MB
llama-server-v1-windows-x64.zip	llama-server for Windows x64, CPU	17 MB

Both llama-server builds are llama.cpp at commit f5222f2 with a small patch that limits tool arguments to the listed choices, so the model can only name places and outputs the robot really has.

SHA-256:

dad3bb39ba0f60e4b4aea4a8b3b87d2350748f65784d9b7424609983a4f95f62  robocor-agent-450m-v1-Q4_0.gguf
ebfc428baa37efad8bae93864f914b2634a09009f91ad59f974fe1a1565d8561  mmproj-robocor-agent-450m-v1-Q8_0.gguf
3e1b6e6054e6591544113ed5f2eeb73d2b6acc8d5fe5c31077f6f07ad23b327a  llama-server-v1-linux-arm64-cuda.tar.gz
1b0c85aa433cfb4bb038b2ec190838b9b33aa86fb58526285d6ffe9abce13190  llama-server-v1-windows-x64.zip

What it does

The model sees a fixed system prompt with the robot's state ("idle, battery 64%.") and up to eleven compact tools: go_to, visit, shuttle, stop, dock, start_patrol, where_am_i, robot_status, take_snapshot, set_output and toggle_output. Place and output arguments are closed lists filled from the robot's saved places and outputs. The app translates each call into the same tool the cloud Agent would use, so operator approvals and the chat transcript work the same way.

It is meant for short commands: trips, stops, docking, patrols, outputs, "where are you", a status check and describing a camera picture. Settings, reports and longer tasks stay with the cloud Agent.

Scores

Exact-match accuracy on Robocor's held-out command sets, with the trained sampling (temperature 0.1, top_k 50, repeat penalty 1.1). Measured on an Apple M-series Mac with stock llama.cpp (no choice-list patch, which adds roughly half a point to two points).

Set	Cases	Accuracy
Main commands	400	98.5%
Unexpected phrasings	77	93.5%
New unexpected phrasings	251	80.9%
Fresh set, written after training	300	81.7%

On a Jetson Orin Nano the Q4_0 language model answers a command in about 0.35 seconds.

Running it yourself

llama-server -m robocor-agent-450m-v1-Q4_0.gguf \
  --mmproj mmproj-robocor-agent-450m-v1-Q8_0.gguf \
  -ngl 99 -c 4096 -ub 256 -np 1 --cache-ram 0 --jinja --reasoning off

The chat template inside the GGUF is the LFM2.5 text template, because llama.cpp cannot build a tool grammar from the original VL template.

License

The model is a derivative of LFM2.5-VL-450M and is distributed under the LFM Open License v1.0. The weights were changed by fine-tuning and the chat template was replaced, as described above. llama.cpp is MIT licensed; the Windows build includes the LLVM OpenMP runtime under its own license (LICENSE-LLVM-OpenMP in the zip).