Total Robots
—
Deployed
—
Provisioned
—
Online (1h)
—
Draft
—
Retired
—
Recent Robots
| Robot ID | Name | Soul | Brain | Status | Actions |
|---|
| Robot ID | Name | Hardware | Soul | Brain | Status | Provisioned | Actions |
|---|
🤖 Identity
✨ Soul
🧠 Brain
⚙️ Hardware
🌐 Network
🔒 Security
When triggered, permanently destroys on-board storage. Only for classified deployments.
Deployed Robots
| Robot ID | Name | Soul | Last Seen | Firmware | Status | Actions |
|---|
📋 Activity Log — All Robots
📡 Push to Single Robot
🌐 Fleet Broadcast
All deployed + provisioned robots
⚠️ This pushes to ALL active robots. Use carefully.
📋 OTA Update History
Loading…
🧬 ATAVUS Genesis-1 R10v2 — Live · Port 8008 · 4.7GB Q4_K_M
ATAVUS Neural Core A1 | Runtime: atavus.cpp v1.0.0 | Fine-tuned for ATAVUS humanoid robotics
Model
ATAVUS Genesis-1 (R10v2) ✅
File
/data/models/genesis-1-gguf/genesis-1-q4_k_m.gguf
Size
4.7 GB (Q4_K_M quantized)
Serve Port
8008 (atavus.cpp, OpenAI-compat)
Training Rounds
10 rounds complete ✅ (best: R10v2)
Final Loss / Acc
0.6436 / 84.2% (R10v2 best)
Training Samples
1,317 (robot: 950 · conversation: 367)
Context
8,192 tokens (deployed) / 40,960 (model max)
ATAVUS Genesis-1 R10v2 — QLoRA fine-tune (rank 128, alpha 256) on 1,317 samples across 10 training rounds. Val loss 0.6436, accuracy 84.2% — best-ever. Covers: manipulation, navigation, sensor fusion, safety/emergency (71), HRI (30), pick-and-place (15), companion conversation (84), multi-turn robot (15+). Via atavus.cpp on port 8008. ● Live since 2026-07-09 12:29 CEST
💬 Live Chat
🏋️ Fine-Tune
📦 Create Update
📊 Dataset
📋 Job Status
🧬 Genesis-1 Live Test
Talk to Genesis-1 — casual conversation, robot commands, sensor tests. It's a real companion, not just a task executor.
Ready — type a message to test Genesis-1...
Quick tests:
Latency:
● Port 8008
1 epoch ≈ 45min CPU. 3 epochs recommended for first run.
What happens when you train:
1. Loads ATAVUS Neural Core A1 base weights
2. Applies QLoRA (rank 64, alpha 128) — precision fine-tuning
3. Fine-tunes on your dataset (215+ samples: language + embodiment)
4. Saves LoRA adapter to
Training History:
R4 (CPU, 215 samples) — 98.5% train acc — overfit, no val split
R5 (GPU, 215 samples) — val loss 2.35, acc 41% — too few samples
R6 (GPU, 378 samples) — val loss 0.875, acc 79.5% — superseded
R7 (GPU, 487 samples) — val loss 0.926, acc 75.8% — regression
R8 (GPU, 959 samples) — val loss 0.659, acc 84.1% — backed up
R9 (GPU, 1,455 samples) — val loss 0.706, acc 83.2% — not deployed
R10v1 (GPU, 1,002 samples) — val loss 0.761 — ❌ failed (wrong dataset)
R10v2 (GPU, 1,317 samples) — val loss 0.6436, acc 84.2% ✓ LIVE 5. Push adapter via OTA to all robots
2. Applies QLoRA (rank 64, alpha 128) — precision fine-tuning
3. Fine-tunes on your dataset (215+ samples: language + embodiment)
4. Saves LoRA adapter to
/data/models/genesis-1/adapter/Training History:
R4 (CPU, 215 samples) — 98.5% train acc — overfit, no val split
R5 (GPU, 215 samples) — val loss 2.35, acc 41% — too few samples
R6 (GPU, 378 samples) — val loss 0.875, acc 79.5% — superseded
R7 (GPU, 487 samples) — val loss 0.926, acc 75.8% — regression
R8 (GPU, 959 samples) — val loss 0.659, acc 84.1% — backed up
R9 (GPU, 1,455 samples) — val loss 0.706, acc 83.2% — not deployed
R10v1 (GPU, 1,002 samples) — val loss 0.761 — ❌ failed (wrong dataset)
R10v2 (GPU, 1,317 samples) — val loss 0.6436, acc 84.2% ✓ LIVE 5. Push adapter via OTA to all robots
⏱️ Est. time per round: ~43 min on RTX 4080 SUPER (vast.ai GPU).
R10v2 live. Next round: R11 — target val loss <0.62 via larger dataset.
R10v2 live. Next round: R11 — target val loss <0.62 via larger dataset.
Build an OTA update package from your current Genesis-1 model config. Deploy to individual robots or broadcast to fleet.
R4 Dataset: 215 samples (115 language + 100 embodiment) · LoRA rank 64 · loss 0.0397 · acc 98.5%
Click Refresh to check
🔌 Prototype Deploy
Pi4 / PiDog prototype provisioning — build, download, and track hardware prototypes
Total Prototypes
—
Active
—
Pending
—
Awaiting first ping
Offline
—
Provisioned Prototypes
| Proto ID | Robot Name | Model | Genesis URL | Status | Deployed | Last Ping | Actions |
|---|---|---|---|---|---|---|---|
| Loading... | |||||||
📋 Pi Setup Guide
🛠 Quick Start
1. Click + New Prototype — configure robot name, persona, Genesis URL
2. Click ⬇ Download Bundle — get a complete ZIP package
3. Copy ZIP to your Pi 4:
scp atavus-pi-*.zip pi@<PI_IP>:~/
4. On your Pi, unzip and run:
unzip atavus-pi-*.zip && cd atavus-pi
sudo bash setup_pi.sh
sudo reboot
sudo bash setup_pi.sh
sudo reboot
5. Pi boots → Chromium opens Atavus UI automatically
⚙️ What's in the Bundle
📄 soul.py — Main AI brain (mic→STT→Genesis-1→TTS→speaker)
📄 body_nerve_pi.py — PiDog sensor daemon (IMU, touch, battery)
📄 soul_pi.json — Your custom config (robot name, persona, Genesis URL)
📄 setup_pi.sh — One-shot installer (whisper.cpp, piper TTS, systemd)
📄 chat_ui/ — Touch-friendly dark web interface
📄 MANIFEST.json — Bundle metadata
✅ Genesis-1 R10v2 running on R6515:8008
✅ Pi acts as I/O layer — brain stays on server
✅ Works offline for UI; needs network for AI responses
✅ Pi acts as I/O layer — brain stays on server
✅ Works offline for UI; needs network for AI responses
🤖 Two Operating Modes
💻 Tabletop Mode
PiDog sits on desk. Chat via touchscreen + voice. Camera watches user. Servos idle. Perfect for testing.
🦾 Humanoid Body Mode
Strip PiDog hardware, embed into humanoid body. Servos active. Body nerve feeds sensors. Genesis-1 sees + feels the body.