Robots
| Name | Robot ID | Status | Brain |
|---|
Fleet devices
| Name | Device ID | Status | Last seen |
|---|
Quick actions
Robots
| Name | Robot ID | Model | Soul | Brain | Status | Firmware | Last seen |
|---|
Fleet health
Devices
| Name | Device ID | Status | Brain | Firmware | Brain ver | Last seen |
|---|
Register update bundle
Bundles
| Version | Kind | Channel | Rollout | Activated | SHA256 |
|---|
Cloud Brain — test offload
Account
- Loading…
Two-Factor Authentication
Change password
Atavus Soul Box — Usage Guide
One console for robots (personality/brain config), the deployed fleet (live devices, updates, safety), and the cloud brain.
Go to Robots → + New robot. Set a name, hardware model, and (optionally) a "soul" (personality, voice, wake word) and "brain" (model, endpoint, system prompt). This defines what the box will run.
Open a robot → Provision. This builds a self-contained .tar.gz (agent + Ed25519 identity + setup). Flash a Jetson AGX Orin (JetPack 7.2) and run setup.sh on first boot — it enrolls the box into the fleet automatically.
A device appears in Fleet once it enrolls (Ed25519-signed). You can also pre-create one under Fleet → + New device and download its bundle from the device detail.
Fleet shows live status (enrolled/offline/revoked), firmware + brain versions, and last heartbeat (with telemetry metrics captured from each heartbeat). Open a device to see its command history and issue actions.
In Updates, register a bundle (version, kind, artifact SHA256), then Activate it. Set Rollout % for staged canary deployment. Devices pull the update over the signed OTA channel and auto-rollback on failure. Sign artifacts with atavus-core/scripts/deployment/artifact.py (Ed25519 + SHA256 + provenance).
Open a device → 🛑 Emergency shutdown. The signed kill command is queued, the device e-stops and powers off on receipt (top priority), and the action is fully audited.
The Cloud Brain tab tests the offload: type a reasoning/planning/critical-problem query and it routes to the cloud model (interim local model now; the fine-tuned Mega Robot Mind later). Devices offload via the signed /brain route; the on-device BrainRouter falls back to the local VLA when offline.
Every head/checkpoint/bundle is SHA256-hashed + Ed25519-signed + provenance-tracked (Atavus Core artifact.py). Devices verify integrity + signature before loading anything — supply-chain security at fleet scale.
Every device holds its own Ed25519 keypair; every message is signed; every bundle/command is server-signed and verified on-device. Revoked devices can never re-authenticate. Legal: Atavus Core v3.0 is an Apache-2.0 derivative of NVIDIA Isaac GR00T N1.7 (attribution retained).
Brain weights (FP16 ~6.3 GB per head) ship via OTA. On the Orin, build TensorRT engines on-device from Atavus Core:
git clone git@github.com:atavusai/atavus-core.git cd atavus-core && bash scripts/deployment/orin/install_deps.sh && source scripts/activate_orin.sh python scripts/deployment/build_trt_pipeline.py --model-path <head> --dataset-path <demo> --embodiment-tag <tag>