jetson/Dockerfile.social

@@ -0,0 +1,119 @@
+# Social-bot container — JetPack 6 + TensorRT + audio AI stack
+# Extends the base ROS2 Humble container with social-bot dependencies.
+#
+# Deps: faster-whisper, llama-cpp-python (CUDA), piper-tts, insightface,
+#       pyannote.audio, OpenWakeWord, Silero VAD, pyaudio, sounddevice
+#
+# Build: docker build -f Dockerfile.social -t saltybot/social:latest .
+# Run:   docker compose -f docker-compose.yml up -d saltybot-social
+
+FROM nvcr.io/nvidia/l4t-jetpack:r36.2.0
+
+LABEL maintainer="sl-jetson"
+LABEL description="Social-bot AI stack — speech, LLM, TTS, face recognition on Orin Nano Super"
+LABEL jetpack="6.0"
+LABEL l4t="r36.2.0"
+
+ENV DEBIAN_FRONTEND=noninteractive
+ENV ROS_DISTRO=humble
+ENV ROS_ROOT=/opt/ros/${ROS_DISTRO}
+ENV PYTHONDONTWRITEBYTECODE=1
+ENV PYTHONUNBUFFERED=1
+
+# ── Locale ────────────────────────────────────────────────────────────────────
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    locales tzdata \
+    && locale-gen en_US.UTF-8 \
+    && rm -rf /var/lib/apt/lists/*
+ENV LANG=en_US.UTF-8
+
+# ── System deps ───────────────────────────────────────────────────────────────
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    build-essential cmake git wget curl ca-certificates \
+    python3-dev python3-pip python3-setuptools python3-wheel \
+    # Audio hardware
+    alsa-utils libasound2-dev pulseaudio pulseaudio-utils \
+    portaudio19-dev libsndfile1-dev libsoundio-dev \
+    # USB audio / ReSpeaker support
+    usbutils libusb-1.0-0-dev \
+    # CUDA / TensorRT tools
+    cuda-toolkit-12-2 libcudnn8 libcudnn8-dev \
+    # Misc
+    htop tmux nano ffmpeg \
+    && rm -rf /var/lib/apt/lists/*
+
+# ── ROS2 Humble ───────────────────────────────────────────────────────────────
+RUN curl -sSL https://raw.githubusercontent.com/ros/rosdistro/master/ros.asc \
+    | gpg --dearmor -o /usr/share/keyrings/ros-archive-keyring.gpg && \
+    echo "deb [arch=arm64 signed-by=/usr/share/keyrings/ros-archive-keyring.gpg] \
+    http://packages.ros.org/ros2/ubuntu jammy main" \
+    > /etc/apt/sources.list.d/ros2.list && \
+    apt-get update && apt-get install -y --no-install-recommends \
+    ros-humble-ros-base \
+    python3-colcon-common-extensions \
+    python3-rosdep \
+    && rm -rf /var/lib/apt/lists/*
+
+# ── Python AI deps (core) ─────────────────────────────────────────────────────
+# faster-whisper: CTranslate2 backend, Orin GPU accelerated
+# llama-cpp-python: GGUF quantized LLM, CUDA offload
+# piper-tts: fast neural TTS, CPU/GPU
+# insightface: SCRFD face detection + ArcFace recognition
+# pyannote.audio: speaker diarization + ECAPA-TDNN embeddings
+RUN pip3 install --no-cache-dir \
+    "faster-whisper>=1.0.0" \
+    "ctranslate2>=4.0.0" \
+    "openai-whisper>=20231117" \
+    "piper-tts>=1.2.0" \
+    "insightface>=0.7.3" \
+    "onnxruntime-gpu>=1.17.0" \
+    "pyannote.audio>=3.1.0" \
+    "speechbrain>=1.0.0"
+
+# ── llama-cpp-python with CUDA ─────────────────────────────────────────────────
+# Build with CUDA support for Orin GPU offload
+RUN CMAKE_ARGS="-DLLAMA_CUDA=on -DCUDA_ARCHITECTURES=87" \
+    pip3 install --no-cache-dir llama-cpp-python==0.2.85 --no-binary llama-cpp-python
+
+# ── Wake word / VAD ───────────────────────────────────────────────────────────
+RUN pip3 install --no-cache-dir \
+    "openwakeword>=0.6.0" \
+    "silero-vad>=5.1.0" \
+    "webrtcvad-wheels>=2.0.14"
+
+# ── Audio I/O ─────────────────────────────────────────────────────────────────
+RUN pip3 install --no-cache-dir \
+    "pyaudio>=0.2.14" \
+    "sounddevice>=0.4.6" \
+    "soundfile>=0.12.1" \
+    "numpy>=1.24.0"
+
+# ── TensorRT Python bindings ──────────────────────────────────────────────────
+# Already available via JetPack; install pycuda for custom kernels
+RUN pip3 install --no-cache-dir "pycuda>=2022.2.2"
+
+# ── MQTT for SOUL/agent communication ─────────────────────────────────────────
+RUN pip3 install --no-cache-dir "paho-mqtt>=2.0.0"
+
+# ── ROS2 Python deps ──────────────────────────────────────────────────────────
+RUN pip3 install --no-cache-dir \
+    "rclpy" \
+    "transforms3d>=0.4.1"
+
+# ── Model directory ───────────────────────────────────────────────────────────
+RUN mkdir -p /models/onnx /models/engines /models/gguf /models/piper
+ENV MODEL_DIR=/models
+ENV PIPER_VOICE_DIR=/models/piper
+
+# ── ALSA config for USB mic + speaker ─────────────────────────────────────────
+COPY config/asound.conf /etc/asound.conf
+
+# ── Workspace ─────────────────────────────────────────────────────────────────
+RUN mkdir -p /ros2_ws/src
+WORKDIR /ros2_ws
+
+COPY scripts/entrypoint.sh /entrypoint.sh
+RUN chmod +x /entrypoint.sh
+
+ENTRYPOINT ["/entrypoint.sh"]
+CMD ["bash"]

jetson/config/asound.conf

@@ -0,0 +1,38 @@
+# ALSA config for social-bot audio
+# Assumes USB mic (ReSpeaker or similar) at card 1, USB speaker at card 2
+# Adjust card numbers with: aplay -l / arecord -l
+
+# Default capture device — USB mic array (ReSpeaker 2-Mic or 4-Mic)
+pcm.!default {
+    type asym
+    playback.pcm "usb_speaker"
+    capture.pcm  "usb_mic"
+}
+
+# USB microphone input
+pcm.usb_mic {
+    type plug
+    slave {
+        pcm "hw:1,0"
+        rate 16000
+        channels 1
+        format S16_LE
+    }
+}
+
+# USB speaker output
+pcm.usb_speaker {
+    type plug
+    slave {
+        pcm "hw:2,0"
+        rate 22050
+        channels 1
+        format S16_LE
+    }
+}
+
+# Mixer controls
+ctl.!default {
+    type hw
+    card 1
+}

jetson/docker-compose.yml

@@ -302,6 +302,61 @@ services:
           fence_active:=true
       "
 
+  # ── Social-bot AI stack (speech + LLM + TTS + face recognition) ─────────────
+  # Issue #88: Orin dev environment for social-bot
+  # Start: docker compose up -d saltybot-social
+  # Logs:  docker compose logs -f saltybot-social
+  saltybot-social:
+    image: saltybot/social:latest
+    build:
+      context: .
+      dockerfile: Dockerfile.social
+    container_name: saltybot-social
+    restart: unless-stopped
+    runtime: nvidia
+    network_mode: host
+    depends_on:
+      - stm32-bridge
+    environment:
+      - NVIDIA_VISIBLE_DEVICES=all
+      - NVIDIA_DRIVER_CAPABILITIES=all,audio
+      - ROS_DOMAIN_ID=42
+      - RMW_IMPLEMENTATION=rmw_cyclonedds_cpp
+      - MODEL_DIR=/models
+      - PIPER_VOICE_DIR=/models/piper
+      - LLAMA_N_GPU_LAYERS=20
+      # Audio device routing (override if mic/speaker cards differ)
+      - ALSA_MIC_CARD=1
+      - ALSA_SPEAKER_CARD=2
+    volumes:
+      - ./ros2_ws/src:/ros2_ws/src:rw
+      - ./config:/config:ro
+      # Persistent model storage on NVMe (survives container restarts)
+      - /mnt/nvme/saltybot/models:/models:rw
+      # Enrollment database (face embeddings, voice prints, conversation history)
+      - /mnt/nvme/saltybot/social_db:/social_db:rw
+      # SOUL.md personality file
+      - /mnt/nvme/saltybot/soul:/soul:ro
+    devices:
+      # USB mic array (ReSpeaker 2-Mic / 4-Mic or compatible)
+      - /dev/snd:/dev/snd
+      # USB webcam / face camera (optional, RealSense preferred)
+      - /dev/bus/usb:/dev/bus/usb
+    group_add:
+      - audio
+    cap_add:
+      - SYS_NICE  # allow real-time thread priority for audio
+    ulimits:
+      rtprio: 95
+      memlock: -1
+    command: >
+      bash -c "
+        source /opt/ros/humble/setup.bash &&
+        source /ros2_ws/install/local_setup.bash 2>/dev/null || true &&
+        ros2 launch saltybot_social social_bot.launch.py
+      "
+
+
 volumes:
   saltybot-maps:
     driver: local

jetson/models/README.md

@@ -0,0 +1,79 @@
+# Social-bot Model Directory
+
+## Layout
+
+```
+/models/
+├── onnx/                          # Source ONNX models (version-pinned)
+│   ├── scrfd_10g_bnkps.onnx      # Face detection — InsightFace SCRFD-10GF
+│   ├── arcface_r100.onnx         # Face recognition — ArcFace R100 (buffalo_l)
+│   └── ecapa_tdnn.onnx           # Speaker embedding — ECAPA-TDNN (SpeechBrain export)
+│
+├── engines/                       # TensorRT FP16 compiled engines
+│   ├── scrfd_10g_fp16.engine     # SCRFD → TRT FP16 (640×640)
+│   ├── arcface_r100_fp16.engine  # ArcFace → TRT FP16 (112×112)
+│   └── ecapa_tdnn_fp16.engine    # ECAPA-TDNN → TRT FP16 (variable len)
+│
+├── whisper-small-ct2/             # faster-whisper CTranslate2 format (auto-downloaded)
+│   ├── model.bin
+│   └── tokenizer.json
+│
+├── piper/                         # Piper TTS voice models
+│   ├── en_US-lessac-medium.onnx
+│   └── en_US-lessac-medium.onnx.json
+│
+├── gguf/                          # Quantized LLM (llama-cpp-python)
+│   └── phi-3-mini-4k-instruct-q4_k_m.gguf  # ~2.2GB — Phi-3-mini Q4_K_M
+│
+└── speechbrain_ecapa/             # SpeechBrain pretrained checkpoint cache
+```
+
+## Model Versions
+
+| Model | Version | Source | Size |
+|---|---|---|---|
+| SCRFD-10GF | InsightFace 0.7 | GitHub releases | 17MB |
+| ArcFace R100 (w600k_r50) | InsightFace buffalo_l | Auto via insightface | 166MB |
+| ECAPA-TDNN | SpeechBrain spkrec-ecapa-voxceleb | HuggingFace | 87MB |
+| Whisper small | faster-whisper 1.0+ | CTranslate2 hub | 488MB |
+| Piper en_US-lessac-medium | Rhasspy piper-voices | HuggingFace | 63MB |
+| Phi-3-mini-4k Q4_K_M | microsoft/Phi-3-mini-4k-instruct | GGUF / HuggingFace | 2.2GB |
+
+## Setup
+
+```bash
+# From within the social container:
+/scripts/convert_models.sh all          # download + convert all models
+/scripts/convert_models.sh benchmark    # run latency benchmark suite
+/scripts/convert_models.sh health       # check GPU memory
+```
+
+## Performance Targets (Orin Nano Super, JetPack 6, FP16)
+
+| Model | Input | Target | Typical |
+|---|---|---|---|
+| SCRFD-10GF | 640×640 | <15ms | ~8ms |
+| ArcFace R100 | 4×112×112 | <5ms | ~3ms |
+| ECAPA-TDNN | 1s audio | <20ms | ~12ms |
+| Whisper small | 1s audio | <300ms | ~180ms |
+| Piper lessac-medium | 10 words | <200ms | ~60ms |
+| Phi-3-mini Q4_K_M | prompt | <500ms TTFT | ~350ms |
+
+## LLM Download
+
+```bash
+# Download Phi-3-mini GGUF manually (2.2GB):
+wget -O /models/gguf/phi-3-mini-4k-instruct-q4_k_m.gguf \
+  "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct-gguf/resolve/main/Phi-3-mini-4k-instruct-q4.gguf"
+
+# Or use llama-cpp-python's built-in download:
+python3 -c "
+from llama_cpp import Llama
+llm = Llama.from_pretrained(
+    repo_id='microsoft/Phi-3-mini-4k-instruct-gguf',
+    filename='Phi-3-mini-4k-instruct-q4.gguf',
+    cache_dir='/models/gguf',
+    n_gpu_layers=20
+)
+"
+```

jetson/scripts/convert_models.sh

@@ -0,0 +1,339 @@
+#!/usr/bin/env bash
+# convert_models.sh — Convert social-bot ONNX models to TensorRT FP16 engines
+#
+# Models converted:
+#   SCRFD-10GF       — face detection (InsightFace)
+#   ArcFace-R100     — face recognition (InsightFace)
+#   Whisper-small    — STT via faster-whisper (handled by CTranslate2, not TRT)
+#   ECAPA-TDNN       — speaker embedding (SpeechBrain export)
+#   Piper TTS        — neural TTS (ONNX native, no TRT needed)
+#
+# Usage: ./scripts/convert_models.sh [--model all|scrfd|arcface|ecapa]
+# Requirements: TensorRT 8.6+, trtexec in PATH, ONNX models in /models/onnx/
+#
+# Note: Whisper uses faster-whisper's CTranslate2 which handles its own GPU
+#       optimization. Piper uses ONNX Runtime directly (already optimized).
+#
+set -euo pipefail
+
+MODEL_DIR="${MODEL_DIR:-/models}"
+ONNX_DIR="${MODEL_DIR}/onnx"
+ENGINE_DIR="${MODEL_DIR}/engines"
+WORKSPACE_MB=4096
+TARGET="${1:-all}"
+
+mkdir -p "${ENGINE_DIR}"
+
+log() { echo "[$(date +%H:%M:%S)] $*"; }
+err() { echo "[ERROR] $*" >&2; exit 1; }
+
+check_trtexec() {
+    command -v trtexec &>/dev/null || \
+        err "trtexec not found. Install TensorRT or add to PATH."
+}
+
+# ── Download ONNX sources ──────────────────────────────────────────────────────
+download_models() {
+    log "Downloading ONNX model sources..."
+
+    # SCRFD-10GF (InsightFace model zoo)
+    SCRFD_ONNX="${ONNX_DIR}/scrfd_10g_bnkps.onnx"
+    if [[ ! -f "${SCRFD_ONNX}" ]]; then
+        log "Downloading SCRFD-10GF..."
+        wget -q --show-progress -O "${SCRFD_ONNX}" \
+            "https://github.com/deepinsight/insightface/releases/download/v0.7/scrfd_10g_bnkps.onnx" || \
+        python3 -c "
+import insightface, shutil, os
+app = insightface.app.FaceAnalysis(providers=['CUDAExecutionProvider'])
+app.prepare(ctx_id=0)
+src = os.path.expanduser('~/.insightface/models/buffalo_l/det_10g.onnx')
+if os.path.exists(src): shutil.copy(src, '${SCRFD_ONNX}')
+"
+    fi
+
+    # ArcFace-R100 (InsightFace buffalo_l)
+    ARCFACE_ONNX="${ONNX_DIR}/arcface_r100.onnx"
+    if [[ ! -f "${ARCFACE_ONNX}" ]]; then
+        log "Downloading ArcFace-R100 via insightface..."
+        python3 -c "
+import insightface, shutil, os
+app = insightface.app.FaceAnalysis(providers=['CUDAExecutionProvider'])
+app.prepare(ctx_id=0)
+src = os.path.expanduser('~/.insightface/models/buffalo_l/w600k_r50.onnx')
+if os.path.exists(src): shutil.copy(src, '${ARCFACE_ONNX}')
+"
+    fi
+
+    # ECAPA-TDNN speaker embedding (SpeechBrain pretrained)
+    ECAPA_ONNX="${ONNX_DIR}/ecapa_tdnn.onnx"
+    if [[ ! -f "${ECAPA_ONNX}" ]]; then
+        log "Exporting ECAPA-TDNN to ONNX..."
+        python3 - <<'PYEOF'
+import torch
+from speechbrain.pretrained import EncoderClassifier
+import os
+
+model = EncoderClassifier.from_hparams(
+    source="speechbrain/spkrec-ecapa-voxceleb",
+    savedir="/models/speechbrain_ecapa"
+)
+model.eval()
+
+dummy = torch.randn(1, 16000)  # 1s of audio
+with torch.no_grad():
+    torch.onnx.export(
+        model.encode_batch.__self__,
+        dummy,
+        "/models/onnx/ecapa_tdnn.onnx",
+        opset_version=12,
+        input_names=["waveform"],
+        output_names=["embedding"],
+        dynamic_axes={"waveform": {1: "audio_len"}}
+    )
+print("ECAPA-TDNN exported to /models/onnx/ecapa_tdnn.onnx")
+PYEOF
+    fi
+
+    log "ONNX models ready in ${ONNX_DIR}"
+}
+
+# ── SCRFD face detection ───────────────────────────────────────────────────────
+convert_scrfd() {
+    local ONNX="${ONNX_DIR}/scrfd_10g_bnkps.onnx"
+    local ENGINE="${ENGINE_DIR}/scrfd_10g_fp16.engine"
+
+    [[ -f "${ONNX}" ]] || err "SCRFD ONNX not found at ${ONNX}. Run with --download first."
+
+    log "Converting SCRFD-10GF → TRT FP16..."
+    trtexec \
+        --onnx="${ONNX}" \
+        --saveEngine="${ENGINE}" \
+        --fp16 \
+        --workspace=${WORKSPACE_MB} \
+        --minShapes=input.1:1x3x640x640 \
+        --optShapes=input.1:1x3x640x640 \
+        --maxShapes=input.1:4x3x640x640 \
+        --explicitBatch \
+        2>&1 | tail -5
+
+    log "SCRFD engine saved: ${ENGINE}"
+    benchmark_engine "${ENGINE}" "input.1:1x3x640x640"
+}
+
+# ── ArcFace recognition ────────────────────────────────────────────────────────
+convert_arcface() {
+    local ONNX="${ONNX_DIR}/arcface_r100.onnx"
+    local ENGINE="${ENGINE_DIR}/arcface_r100_fp16.engine"
+
+    [[ -f "${ONNX}" ]] || err "ArcFace ONNX not found at ${ONNX}."
+
+    log "Converting ArcFace-R100 → TRT FP16..."
+    trtexec \
+        --onnx="${ONNX}" \
+        --saveEngine="${ENGINE}" \
+        --fp16 \
+        --workspace=${WORKSPACE_MB} \
+        --minShapes=input.1:1x3x112x112 \
+        --optShapes=input.1:4x3x112x112 \
+        --maxShapes=input.1:16x3x112x112 \
+        --explicitBatch \
+        2>&1 | tail -5
+
+    log "ArcFace engine saved: ${ENGINE}"
+    benchmark_engine "${ENGINE}" "input.1:4x3x112x112"
+}
+
+# ── ECAPA-TDNN speaker embedding ───────────────────────────────────────────────
+convert_ecapa() {
+    local ONNX="${ONNX_DIR}/ecapa_tdnn.onnx"
+    local ENGINE="${ENGINE_DIR}/ecapa_tdnn_fp16.engine"
+
+    [[ -f "${ONNX}" ]] || err "ECAPA-TDNN ONNX not found at ${ONNX}."
+
+    log "Converting ECAPA-TDNN → TRT FP16..."
+    trtexec \
+        --onnx="${ONNX}" \
+        --saveEngine="${ENGINE}" \
+        --fp16 \
+        --workspace=${WORKSPACE_MB} \
+        --minShapes=waveform:1x8000 \
+        --optShapes=waveform:1x16000 \
+        --maxShapes=waveform:1x32000 \
+        --explicitBatch \
+        2>&1 | tail -5
+
+    log "ECAPA-TDNN engine saved: ${ENGINE}"
+    benchmark_engine "${ENGINE}" "waveform:1x16000"
+}
+
+# ── Whisper STT (CTranslate2 — no TRT conversion needed) ─────────────────────
+setup_whisper() {
+    log "Setting up faster-whisper (CTranslate2 format)..."
+    local WHISPER_DIR="/models/whisper-small-ct2"
+
+    if [[ ! -d "${WHISPER_DIR}" ]]; then
+        python3 - <<'PYEOF'
+from faster_whisper import WhisperModel
+import os
+
+# Download and cache the small model in CTranslate2 int8 format
+# This is already GPU-optimized by CTranslate2 — no TRT conversion needed
+model = WhisperModel("small", device="cuda", compute_type="float16",
+                     download_root="/models")
+print("faster-whisper 'small' model ready at /models/")
+# Run one warmup inference
+import numpy as np
+audio = np.zeros(16000, dtype=np.float32)
+segments, _ = model.transcribe(audio)
+list(segments)  # consume generator
+print("Whisper warmup complete")
+PYEOF
+    else
+        log "Whisper already downloaded, skipping."
+    fi
+}
+
+# ── Piper TTS (ONNX Runtime — no TRT conversion needed) ──────────────────────
+setup_piper() {
+    log "Setting up Piper TTS voice models..."
+    local PIPER_DIR="${MODEL_DIR}/piper"
+    mkdir -p "${PIPER_DIR}"
+
+    # Download en_US-lessac-medium (natural, good quality, ~60ms on Orin)
+    if [[ ! -f "${PIPER_DIR}/en_US-lessac-medium.onnx" ]]; then
+        python3 - <<'PYEOF'
+import urllib.request, os
+
+base = "https://huggingface.co/rhasspy/piper-voices/resolve/main/en/en_US/lessac/medium/"
+piper_dir = "/models/piper"
+
+for fname in ["en_US-lessac-medium.onnx", "en_US-lessac-medium.onnx.json"]:
+    dest = os.path.join(piper_dir, fname)
+    if not os.path.exists(dest):
+        print(f"Downloading {fname}...")
+        urllib.request.urlretrieve(base + fname, dest)
+        print(f"  → {dest}")
+print("Piper voice ready.")
+PYEOF
+    else
+        log "Piper voice already downloaded, skipping."
+    fi
+}
+
+# ── Benchmark helper ───────────────────────────────────────────────────────────
+benchmark_engine() {
+    local ENGINE="$1"
+    local SHAPE="$2"
+    log "Benchmarking ${ENGINE##*/}..."
+    trtexec \
+        --loadEngine="${ENGINE}" \
+        --shapes="${SHAPE}" \
+        --warmUp=3 \
+        --avgRuns=20 \
+        --useSpinWait \
+        2>&1 | grep -E "mean|median|throughput|Latency" | head -5
+}
+
+# ── Full benchmark suite ───────────────────────────────────────────────────────
+run_benchmarks() {
+    log "=== Social-bot model benchmark suite ==="
+    log "Platform: $(cat /proc/device-tree/model 2>/dev/null || echo 'Orin')"
+    log "TensorRT: $(python3 -c 'import tensorrt as trt; print(trt.__version__)' 2>/dev/null)"
+    log "CUDA: $(nvcc --version 2>/dev/null | grep release | awk '{print $6}' | tr -d ,)"
+    echo
+
+    for ENGINE in "${ENGINE_DIR}"/*.engine; do
+        [[ -f "${ENGINE}" ]] || continue
+        NAME="${ENGINE##*/}"
+        case "${NAME}" in
+            scrfd*)  SHAPE="input.1:1x3x640x640" ;;
+            arcface*)SHAPE="input.1:4x3x112x112" ;;
+            ecapa*)  SHAPE="waveform:1x16000" ;;
+            *)       SHAPE="" ;;
+        esac
+        [[ -z "${SHAPE}" ]] && continue
+        echo "── ${NAME} ──"
+        benchmark_engine "${ENGINE}" "${SHAPE}"
+        echo
+    done
+
+    log "Whisper (CTranslate2 GPU):"
+    python3 - <<'PYEOF'
+import time, numpy as np
+from faster_whisper import WhisperModel
+model = WhisperModel("small", device="cuda", compute_type="float16", download_root="/models")
+audio = np.random.randn(16000).astype(np.float32)
+# warmup
+for _ in range(3):
+    list(model.transcribe(audio)[0])
+# benchmark
+times = []
+for _ in range(10):
+    t = time.perf_counter()
+    list(model.transcribe(audio)[0])
+    times.append((time.perf_counter() - t) * 1000)
+print(f"  1s audio: mean={sum(times)/len(times):.1f}ms, min={min(times):.1f}ms")
+PYEOF
+
+    log "Piper TTS (ONNX Runtime):"
+    python3 - <<'PYEOF'
+import time
+from piper import PiperVoice
+voice = PiperVoice.load("/models/piper/en_US-lessac-medium.onnx")
+text = "Hello, I am Salty, your personal robot companion."
+# warmup
+for _ in range(3):
+    list(voice.synthesize_stream_raw(text))
+# benchmark
+times = []
+for _ in range(10):
+    t = time.perf_counter()
+    audio_bytes = b"".join(voice.synthesize_stream_raw(text))
+    times.append((time.perf_counter() - t) * 1000)
+print(f"  '{text[:30]}...': mean={sum(times)/len(times):.1f}ms, min={min(times):.1f}ms")
+PYEOF
+}
+
+# ── tegrastats snapshot ────────────────────────────────────────────────────────
+check_gpu_health() {
+    log "GPU / system health check:"
+    tegrastats --interval 1000 &
+    TPID=$!
+    sleep 2
+    kill ${TPID} 2>/dev/null
+    echo
+    log "GPU memory:"
+    nvidia-smi --query-gpu=memory.used,memory.free,memory.total,utilization.gpu \
+        --format=csv,noheader,nounits 2>/dev/null || \
+    python3 -c "import pycuda.driver as drv; drv.init(); d=drv.Device(0); c=d.make_context(); print('Free:', drv.mem_get_info()[0]//1024//1024, 'MB'); c.pop()"
+}
+
+# ── Main ───────────────────────────────────────────────────────────────────────
+check_trtexec
+
+case "${TARGET}" in
+    all)
+        download_models
+        convert_scrfd
+        convert_arcface
+        convert_ecapa
+        setup_whisper
+        setup_piper
+        check_gpu_health
+        run_benchmarks
+        ;;
+    scrfd)    download_models; convert_scrfd ;;
+    arcface)  download_models; convert_arcface ;;
+    ecapa)    download_models; convert_ecapa ;;
+    whisper)  setup_whisper ;;
+    piper)    setup_piper ;;
+    benchmark)run_benchmarks ;;
+    health)   check_gpu_health ;;
+    download) download_models ;;
+    *)
+        echo "Usage: $0 [all|scrfd|arcface|ecapa|whisper|piper|benchmark|health|download]"
+        exit 1
+        ;;
+esac
+
+log "Done."

jetson/systemd/install_systemd.sh

@@ -0,0 +1,33 @@
+#!/usr/bin/env bash
+# install_systemd.sh — Install saltybot systemd services on Orin
+# Run as root: sudo ./systemd/install_systemd.sh
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+REPO_DIR="$(dirname "${SCRIPT_DIR}")"
+SYSTEMD_DIR="/etc/systemd/system"
+DEPLOY_DIR="/opt/saltybot/jetson"
+
+log() { echo "[install_systemd] $*"; }
+
+[[ "$(id -u)" == "0" ]] || { echo "Run as root"; exit 1; }
+
+# Deploy repo to /opt/saltybot/jetson
+log "Deploying to ${DEPLOY_DIR}..."
+mkdir -p "${DEPLOY_DIR}"
+rsync -a --exclude='.git' --exclude='__pycache__' \
+    "${REPO_DIR}/" "${DEPLOY_DIR}/"
+
+# Install service files
+log "Installing systemd units..."
+cp "${SCRIPT_DIR}/saltybot.target"        "${SYSTEMD_DIR}/"
+cp "${SCRIPT_DIR}/saltybot-social.service" "${SYSTEMD_DIR}/"
+
+# Reload and enable
+systemctl daemon-reload
+systemctl enable saltybot.target
+systemctl enable saltybot-social.service
+
+log "Services installed. Start with:"
+log "  systemctl start saltybot-social"
+log "  journalctl -fu saltybot-social"

jetson/systemd/saltybot-social.service

@@ -0,0 +1,32 @@
+[Unit]
+Description=Saltybot Social-Bot Stack (speech + LLM + TTS + face recognition)
+Documentation=https://gitea.vayrette.com/seb/saltylab-firmware
+After=docker.service network-online.target sound.target
+Requires=docker.service
+PartOf=saltybot.target
+
+[Service]
+Type=simple
+Restart=on-failure
+RestartSec=10s
+TimeoutStartSec=120s
+TimeoutStopSec=30s
+
+User=root
+WorkingDirectory=/opt/saltybot/jetson
+
+# Pull latest image before start (optional — comment out for air-gapped deploy)
+ExecStartPre=-/usr/bin/docker compose -f docker-compose.yml pull saltybot-social
+
+# Start only the social service (not the entire stack)
+ExecStart=/usr/bin/docker compose -f docker-compose.yml up --no-recreate saltybot-social
+
+ExecStop=/usr/bin/docker compose -f docker-compose.yml stop saltybot-social
+
+# Logging
+StandardOutput=journal
+StandardError=journal
+SyslogIdentifier=saltybot-social
+
+[Install]
+WantedBy=saltybot.target multi-user.target

jetson/systemd/saltybot.target

@@ -0,0 +1,8 @@
+[Unit]
+Description=Saltybot Full Stack Target
+Documentation=https://gitea.vayrette.com/seb/saltylab-firmware
+After=docker.service network-online.target
+Requires=docker.service
+
+[Install]
+WantedBy=multi-user.target