TY1200 Platform — ROSClaw Platform Pack for the Iluvatar TY1200

中文文档

A complete ROSClaw platform-operations package for the Iluvatar TY1200 embodied-AI compute box (Intel Core Ultra 7 255H + Iluvatar GPGPU MXM, 16/32 GB HBM2e). It turns the vendor's manual docker exec model-serving setup into a controlled, auditable, agent-operable stack.

What's inside

Asset	Description
rosclaw_ty1200 Python package	Platform detection, ixsmi telemetry, evidence bundles, modeld controlled-ops daemon, OpenAI-compatible provider runtime, executable skill runner
skills/ty1200-platform-ops	Platform Operations Skill (declarative assets + executable runner)
providers/	4 provider assets: 2 local vLLM services + 2 remote LLM fallbacks
deployment/	Docker Compose (digest-pinned), systemd units, installer
scripts/	Ops entry points (preflight / verify / start / stop / support-bundle)

The two local model services

Port	Service	Capabilities
8000	Qwen3-Embedding-0.6B	embedding.text, memory.embed, retrieval.embed_*
8001	Cosmos-Reason2-2B	vlm.physical_reasoning, vlm.video_question_answering, critic.* (advisory only — never drives robots directly)

Measured on TY1200 (16 GB): embedding 893 texts/s @batch128, Chinese retrieval top1 3/3; Cosmos TTFT 36 ms, 363 tok/s, 100% success over 20-request runs; 8 s video reasoning in 6.3 s end to end.

Hardware coverage (read-only)

The skill inventories the whole box, not just the model stack:

Class	Detected on this unit
Layered compute	GPGPU (MRC-V100) + NPU (intel_vpu, /dev/accel/accel0) + Arc iGPU (renderD128) + P/E-core topology
EtherCAT	2× Intel igb RJ45 (enp5s0/enp6s0) — link state reported
CAN-FD	PCI CANBUS controller detected; hardware_present_no_driver surfaced honestly
Serial	RS232/485 UARTs (real-IRQ filtered)
I2C / SPI / GPIO	bus/adapters enumerated; unexposed GPIO/SPI reported as WARN, never silently skipped
Display / wireless	HDMI connector state, Wi-Fi (rtw89), BT (hci0), 5G slot, ATSHA204A secure chip

Ops: peripheral_inventory, accelerator_check, rt_check (PREEMPT_RT assessment). Boundary: GPIO writes, CAN/EtherCAT frames, serial/I2C/SPI transactions and any robot motion are robot-domain operations and are explicitly out of scope for this skill.

Architecture

Agent
  ↓
ty1200-platform-ops Skill  (ty1200-ops CLI / SkillRunner)
  ↓  Unix socket, fixed allowlist — no arbitrary shell/docker
rosclaw-ty1200-modeld  (systemd, unprivileged service account)
  ↓  compose file hash-pinned, image digests pinned, approval tokens sha256-checked
Docker Compose model stack
  ↓
vLLM :8000 (embedding)  /  :8001 (cosmos, video via read-only Media Store)
  ↑
Provider Router — capability routing with remote fallback;
media inputs NEVER leave the box

Quick start

# 1. install (setuptools < 61 targets need the shim; no PyPI required)
pip install --no-build-isolation ./ty1200-platform

# 2. preflight & verify live services
./scripts/ty1200-preflight --json
./scripts/ty1200-model-verify --json --with-inference

# 3. full install: modeld + systemd + compose (requires sudo)
sudo ./deployment/install.sh

# 4. use the skill
ty1200-ops '{"operation":"doctor"}'
ty1200-ops '{"operation":"provider_verify","target_service":"embedding"}'
ty1200-ops '{"operation":"model_stack_restart","target_service":"cosmos","approval_token":"<token>"}'

Security model (L0–L6)

• L0/L1 (inspect, doctor, verify, benchmark) — agent may run directly
• L2 (start/stop/restart) — one-shot approval token, verified by modeld as sha256
• L3 (image pull) — only digest-pinned images from the hash-pinned compose file
• L4+ (compose changes, drivers, kernel, firmware) — humans only, outside the skill
• Cosmos output is advisory (executable: false); motion stays behind ROSClaw Action/Permit/Lease/Sandbox
• Every execution writes an evidence bundle (manifest + sha256); tokens, credentials and chain-of-thought are redacted (hash-only)

Tested on

• TY1200, Ubuntu 22.04, CoreX SDK 4.4.0, Docker 29, Python 3.10
• 70 unit/integration/fault-injection/acceptance tests + 13 skill black-box tests
• Live fault injection: docker kill → honest unhealthy → modeld recovery in ~70 s
• Dojo evaluation: ty1200-dojo executes the declared scenarios (8/8 pass on-device)
• 24h availability: rosclaw-ty1200-health.timer samples every 5 min; ty1200-availability --hours 24 evaluates the ≥99.5% gate

Docs

• Peripheral enablement runbook — CAN driver, EtherCAT (IgH/SOEM), GPIO pinctrl, SPI, PREEMPT_RT, 5G, NPU/OpenVINO, ATSHA204A

Notes for deployment at your site

• providers/deepseekv4 and providers/qwen3.6-27b contain site-specific endpoints (10.10.217.108:*) used by the origin deployment — replace with your own OpenAI-compatible services or delete these two provider packs.
• Image references point at the vendor Harbor registry; digests are pinned in deployment/env/ty1200-models.env.example.
• Model weights are not distributed (Cosmos: NVIDIA Open Model License; Qwen3-Embedding: Apache-2.0).

License

MIT (code). Model weights are covered by their own licenses and are not redistributed.