feat: release 4.0 (npu workflow)

.dockerignore

@@ -0,0 +1,23 @@
+# Build artefacts — contain hardcoded host paths (CMakeCache.txt etc.)
+sim/build/
+sim/user/
+ros2_ws/build/
+ros2_ws/install/
+ros2_ws/log/
+
+# Cloned at image-build time by pixi tasks — exclude to keep context small
+gems/
+external/
+inference/
+
+# Pixi environment cache
+.pixi/
+
+# Large model files
+models/
+
+# Developer files
+.vscode/
+.env
+*.egg-info/
+__pycache__/

.github/workflows/build_and_push_docker.yaml

@@ -4,52 +4,45 @@ on:
   push:
     branches:
       - main
-      - development
   workflow_dispatch:
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   build-and-push:
-    runs-on: self-hosted
+    runs-on: [self-hosted, cpu, heavy, gpu]
     steps:
       - name: Checkout code
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
+        with:
+          lfs: true
 
       - name: Log in to Docker Hub
         uses: docker/login-action@v3
         with:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_TOKEN }}
 
-      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v3
+      - name: Build and push all images
+        run: |
+          docker compose -f docker/compose.yaml build
+          docker compose -f docker/compose.yaml push
 
-      - name: Build and push ROS 2 image
-        uses: docker/build-push-action@v6
-        with:
-          context: .
-          file: docker/Dockerfile.ros2
-          push: true
-          tags: |
-            robotecai/mobile-manipulator-demo-ros2${{ github.ref_name == 'development' && '-development' || '' }}:latest
-
-      - name: Build and push Agents image
-        uses: docker/build-push-action@v6
+  smoke-test:
+    needs: build-and-push
+    runs-on: [self-hosted, cpu, heavy, gpu]
+    steps:
+      - name: Log in to Docker Hub
+        uses: docker/login-action@v3
         with:
-          context: .
-          file: docker/Dockerfile.agents
-          push: true
-          build-args: |
-            ROS2_IMAGE=robotecai/mobile-manipulator-demo-ros2${{ github.ref_name == 'development' && '-development' || '' }}:latest
-          tags: |
-            robotecai/mobile-manipulator-demo-agents${{ github.ref_name == 'development' && '-development' || '' }}:latest
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_TOKEN }}
 
-      - name: Build and push O3DE image
-        uses: docker/build-push-action@v6
-        with:
-          context: .
-          file: docker/Dockerfile.o3de
-          push: true
-          build-args: |
-            ROS2_IMAGE=robotecai/mobile-manipulator-demo-ros2${{ github.ref_name == 'development' && '-development' || '' }}:latest
-          tags: |
-            robotecai/mobile-manipulator-demo-o3de${{ github.ref_name == 'development' && '-development' || '' }}:latest
+      - name: Headless sim smoke test (/clock)
+        run: |
+          IMAGE="robotecai/agentic-mobile-manipulator-o3de:latest"
+          docker pull "$IMAGE"
+          docker run --rm --entrypoint pixi "$IMAGE" run sim-smoke
+

.gitignore

@@ -1,6 +1,6 @@
 gems/
 engine/
-inference/
+external/
 ros2_ws/install
 ros2_ws/build
 ros2_ws/src/packages
@@ -19,3 +19,8 @@ safety_violations.json
 imgui.ini
 anomaly_images/
 package_condition_images/
+models/
+runs/
+*.bak0
+sim/user/
+backtrace.log

.gitmodules

@@ -0,0 +1,8 @@
+[submodule "inference/llama.cpp"]
+	path = inference/llama.cpp
+	url = https://github.com/ggml-org/llama.cpp.git
+	shallow = true
+[submodule "inference/FastFlowLM"]
+	path = inference/FastFlowLM
+	url = https://github.com/RobotecAI/FastFlowLM.git
+	shallow = true

README.md

@@ -5,6 +5,11 @@ _Physical AI for Warehouse Robotics_
 <div align="center">
 
 ![Small warehouse featuring Robotnik Kairos](docs/images/demo_comp.jpg)
+
+![CPU: ready](https://img.shields.io/badge/CPU-ready-ED1C24?style=for-the-badge&logo=amd&logoColor=white&labelColor=000000)
+![GPU: ready](https://img.shields.io/badge/GPU-ready-ED1C24?style=for-the-badge&logo=amd&logoColor=white&labelColor=000000)
+![NPU: ready](https://img.shields.io/badge/NPU-ready-ED1C24?style=for-the-badge&logo=amd&logoColor=white&labelColor=000000)
+
 ![Static Badge](https://img.shields.io/badge/AgenticAI-darkred)
 ![Static Badge](https://img.shields.io/badge/EmbodiedAI-black)
 ![Static Badge](https://img.shields.io/badge/LLM-darkred)
@@ -41,6 +46,7 @@ Backup video link: https://vimeo.com/1138627688/dbfa881a27
 - Hardware platform: **AMD Ryzen™ AI** processor hosting robotics stack and embodied agentic AI multi-agent system
 - Model stack: [**Liquid AI's LFM2-VL**](https://www.liquid.ai/) (Vision-Language Model) optimized for AMD hardware
 - Local LLM (gpt-oss-20b) for planning, instruction parsing, and dialogue
+- Local inference across **CPU, GPU (llama.cpp / Vulkan) and the AMD Ryzen™ AI NPU (FastFlowLM)** — the safety VLM runs on the NPU, freeing the GPU for the LLM and perception VLM
 - Robot platform: Mobile manipulator (Robotnik Kairos) in ROS 2 simulated warehouse environment
 - Simulation setup: Hardware-in-the-loop mirrors real-world deployment for safe testing
 - Benefits: On-device reasoning, low latency, privacy, and seamless simulation-to-hardware transition

cloud_config.toml

@@ -1,51 +1,88 @@
+# ════════════════════════════════════════════════════════════════════════════
+#  Cloud variant of the inference SSOT (see config.toml for the full schema).
+#
+#  Here the chat/vision endpoints are remote (backend = "openai") and therefore
+#  NOT served locally — `pixi run inference` skips backend = "openai". Only the
+#  RAG embedding/reranker still run locally on llama.cpp, exactly as in
+#  config.toml.
+#
+#  To use it: everything (agents, `pixi run inference`, download-models) reads
+#  config.toml, so swap this file in and provide the API key:
+#      cp cloud_config.toml config.toml && export OPENAI_API_KEY=sk-...
+#
+#  Because cloud agents use several distinct remote models, endpoints are split
+#  per model (chat_mini / vlm_mini / vlm_nano) and agents reference the right one.
+# ════════════════════════════════════════════════════════════════════════════
+
+# ─── Endpoints ──────────────────────────────────────────────────────────────
+
+[endpoints.chat_mini]
+type     = "llm"
+backend  = "openai"
+model    = "gpt-5-mini"
+base_url = "https://api.openai.com/v1"
+
+[endpoints.vlm_mini]
+type     = "vlm"
+backend  = "openai"
+model    = "gpt-5-mini"
+base_url = "https://api.openai.com/v1"
+
+[endpoints.vlm_nano]
+type     = "vlm"
+backend  = "openai"
+model    = "gpt-5-nano"
+base_url = "https://api.openai.com/v1"
+
+# RAG models are the only endpoints still served locally (backend = "gpu"):
+# `pixi run inference` launches these two llama.cpp servers and skips the
+# remote backend = "openai" endpoints above. Same specs as config.toml.
+[endpoints.embeddings]
+type       = "embedding"
+backend    = "gpu"
+model      = "qwen3-embedding:0.6b"
+host       = "localhost"
+port       = 8082
+model_path = "${DEMO_ROOT}/models/Qwen3-Embedding-0.6B-Q8_0.gguf"
+model_url  = "https://huggingface.co/Qwen/Qwen3-Embedding-0.6B-GGUF/resolve/main/Qwen3-Embedding-0.6B-Q8_0.gguf?download=true"
+
+# NOTE: The reranker and the embedding model are two separate models!
+[endpoints.reranker]
+type       = "reranker"
+backend    = "gpu"
+model      = "qwen3-reranker:0.6b"
+host       = "localhost"
+port       = 8083
+model_path = "${DEMO_ROOT}/models/qwen3-reranker-0.6b-q8_0.gguf"
+model_url  = "https://huggingface.co/ggml-org/Qwen3-Reranker-0.6B-Q8_0-GGUF/resolve/main/qwen3-reranker-0.6b-q8_0.gguf?download=true"
+
+# ─── Agents ─────────────────────────────────────────────────────────────────
+
 [general]
-llm_model = "gpt-5-mini"
-llm_base_url = "https://api.openai.com/v1"
+llm = "chat_mini"
+vlm = "vlm_mini"
 llm_reasoning = false
-vlm_model = "gpt-5-mini"
-vlm_base_url = "https://api.openai.com/v1"
 vlm_reasoning = false
 
-# megamind agent utilizes llm for task delegation
-# to agents using both llm and vlm for task execution
 [megamind_agent]
-llm_model = "gpt-5-mini"
-llm_base_url = "https://api.openai.com/v1"
-vlm_model = "gpt-5-mini"
-vlm_base_url = "https://api.openai.com/v1"
+llm = "chat_mini"
+vlm = "vlm_mini"
 llm_reasoning = false
 vlm_reasoning = false
 
-# inspection agent uses vlm to determine anomalies in the image
 [inspection_agent]
-vlm_model = "gpt-5-nano"
-vlm_base_url = "https://api.openai.com/v1"
+vlm = "vlm_nano"
 vlm_reasoning = false
 vlm_temperature = 0.0
 
 [condition_agent]
-vlm_model = "gpt-5-mini"
-vlm_base_url = "https://api.openai.com/v1"
+vlm = "vlm_mini"
 vlm_reasoning = false
 vlm_temperature = 0.1
 
-# safety agent uses vlm to determine anomalies in the image
-# and embeddings to determine OSHA safety violations
 [safety_agent]
-vlm_model = "gpt-5-nano"
-vlm_base_url = "https://api.openai.com/v1"
+vlm = "vlm_nano"
 vlm_reasoning = false
 vlm_temperature = 0.0
-
-embeddings_model = "qwen3-embedding:0.6b"
-embeddings_base_url = "http://safety-embeddings:8082/v1"
-# NOTE: The reranker and the embedding model are two separate models!
-reranker_base_url = "http://safety-reranker:8083/v1/reranking"
-
-# INFO:
-# Start the EMBEDDING model in llama.cpp using the following command:
-# ./llama-server -m /home/user/llamas/gguf/custom_ggufs/Qwen3-Embedding-0.6b_Q8_0.gguf --embedding --pooling last -c 4096 -b 2048 -ub 2048 --port 8082
-
-# INFO:
-# Start the RERANKER model in llama.cpp using the following command:
-# ./llama-server -m /home/user/llamas/gguf/custom_ggufs/Qwen3-Reranker-0.6B_BF16.gguf --embedding --pooling rank -c 4096  -b 2048 -ub 2048 --port 8083
+embeddings = "embeddings"
+reranker = "reranker"