Where Robot Services Have Landed × Software Worth Borrowing/Integrating — 6 Deployed Domains × 5-Layer Software Stack (Open vs Closed) × Where We Can Build (Heterogeneous Fleet Middleware / Marsupial Coordination / Data & Eval Tools, all links verified)

⚠️ Research snapshot (2026-07-01). Answers the boss's two questions: ① which domains have actually deployed robot services, and ② what software they use — where can a small software/integration team play? Starting point: 7 Phoenix (ifeng) reports (robot school / JD JoyRobocare / AgiBot going to Europe / world models / UBTECH U1 / Apptronik school / AgiBot G2 in a factory). Every key figure and license detail carries an external [source] link; uncertain points are flagged, not invented. Sister pieces: Coordination-software landscape · SEER architecture teardown · Other embodied-AI atlas.

🧭 The one-line takeaway

Robot bodies (hardware) and base models are already locked up by big vendors + China's "national team" with heavy capital, so a small team should not build a body, nor train a base model from scratch. The positions that are still friendly to a software team — and still a "narrow blue ocean" — are four: ① multi-vendor heterogeneous fleet orchestration middleware (incl. the "marsupial coordination" gap that has no common protocol) ② robot data curation/QC/eval SaaS ③ VLA "eval-as-a-service + edge inference" middleware ④ vertical last-mile system integration. All four share one trait — no hardware, no compute/data moat — and harvest the dividend of "fragmented ecosystem, unsettled standards." Of these, ① fits our existing "marsupial coordination software + SEER reference" work best and should be the main thrust.

🏭 1. Which domains have deployed robot services (all with real cases + scale)

🧩 2. What software these services use (5-layer stack + open/closed)

① Brain (VLA base models)

• Free-to-use open: π0 / π0.5 (openpi, Apache-2.0 code; weights under Gemma terms; inference needs only >8GB / a 4090, LoRA on a single card)¹⁶; NVIDIA GR00T N1→N1.7 (Apache code; weights commercial from N1.7)¹⁷; LeRobot / SmolVLA (450M small model, runs on consumer hardware)¹⁸.
• China open (all CC BY-NC, non-commercial): AgiBot GO-1¹⁹, Galaxea G0²⁰, X Square WALL-OSS (already in the LeRobot ecosystem)²¹.
• Closed: Tesla Optimus, Figure Helix, Skild Brain, Google Gemini Robotics (hybrid API/trusted-tester), Noematrix, Galbot GroceryVLA.

② Simulation & world models (the sim2real base): NVIDIA Isaac Sim/Lab (Apache/BSD open) + Cosmos world models (open, commercial-OK)²²; MuJoCo/Playground (Apache open); Genesis (Apache open, though its "43M FPS" was shown to be ~150× overstated).

③ Data collection / teleoperation (the bottleneck layer): open hardware ALOHA/UMI/GELLO (mostly MIT — phone/VR/handheld grippers enable crowdsourced capture); data firms Scale AI, Encord; JD's first "embodied-AI data exchange" (600k people, 10M hours over 2 years).

④ Multi-robot orchestration middleware (the "OS" layer of deployment): standards VDA5050 (the "command" layer), MassRobotics (the "monitor" layer), open Open-RMF / openTCS (see the sister piece)²⁴; commercial SEER RDS, Hikrobot RCS, Geek+ RMS, InOrbit (dual-standard native), Viam²⁶, Intrinsic Flowstate (folded into Google in 2026)²⁷.

⑤ Service / ops cloud (RaaS SaaS): Keenon's scheduling platform (+ KONE elevator API), Pudu PuduOS/Pudu Cloud, Gausium FieldBots (>2,000 robots managed, already wired to Open-RMF + InOrbit)¹², Agility Arc²⁵, Boston Dynamics Orbit — each a closed ecosystem of its own.

🎯 3. Where WE can build software / integrate (the core)

🔴 Avoid (red ocean / high moat)

• Robot bodies, VLA base models (capital/data/compute moats are enormous; the national team + big vendors are fighting here).
• Single-vendor fleet management (FMS) — standard on every AGV/AMR vendor; pure-software independents are shaking out (Rocos acquired by Boston Dynamics, Freedom/Ready Robotics shut down).
• Plain LoRA fine-tuning scripts (LeRobot/GR00T/openpi ship these free).

🟢 Worth it (narrow blue ocean, ranked by fit)

1. Multi-vendor heterogeneous fleet orchestration middleware (best fit with our marsupial work)

• Pain: VDA5050 / MassRobotics / Open-RMF coexist and aren't unified; China's vendors (Hikrobot/Geek+/SEER) are "own-fleet first," so A's AMR + B's AGV + C's goods-to-person can't run under one system.
• Entry: a neutral VDA5050 gateway + cross-standard unified orchestration + WMS↔RCS bridge; use Open-RMF/openTCS to lower the start cost. Pitch = "not locked to one brand" (use Zebra winding down Fetch as the cautionary tale)⁸. Market CAGR ~20.9%.
• Marsupial / mother-child shuttle coordination: deployed in AS/RS but with no common protocol, each vendor closed — a differentiation gap that continues our "marsupial coordination software" work.
• Key proof point: Gausium voluntarily wired itself into Open-RMF (official fleet_adapter_ecobot)¹³and the InOrbit connector¹⁴— proof that "multi-brand robots in one pool" is real, the hooks exist, and leading vendors want to be plugged in.

2. Robot data curation / QC / eval SaaS

• "Data, not the model" is the acknowledged bottleneck. Mature commercial tools for cleaning/dedup/diversity-scoring/auto-QC are missing.
• Multimodal gap: existing datasets are mostly vision-only — force/tactile/proprioception aligned with vision is a clear technical gap.

3. VLA "eval-as-a-service + edge inference" middleware (a software team's sweet spot)

• The official eval stack is fragmented and interfaces haven't converged. Build an "any-VLA × any-benchmark × any-robot" unified eval-serve gateway (a vLLM-for-robots).
• Edge gap: real control needs 50–100Hz, but the fastest VLA on a Jetson Orin only reaches 10–20Hz — quantization/compilation pipelines are in demand.

4. Vertical last-mile system integration / multi-brand ops

• Multi-brand ops for service robots, elevator/IoT integration, inspection AI defect-detection SaaS, analytics dashboards.
• The software backbone for a JoyRobocare-style "install—repair—recycle" service network. Fastest cash flow.

Threshold notes

• The dirtiest, heaviest moat of a fleet middleware = writing and maintaining many Fleet Adapters (each vendor's private API + VDA5050/MassRobotics) + real-time traffic deconfliction + elevator/door/charging resource management + standard neutrality.
• Data/eval directions are friendliest to a pure-software team (no hardware/compute), but real-robot closed-loop validation is the hidden threshold — you need real robots or high-fidelity sim.

🔗 4. How this ties into our existing work

• Our "marsupial coordination software (in-house vs bolt-on hardware)" and "SEER feasibility / in-house reference" both sit squarely in 🟢1 (heterogeneous fleet middleware + marsupial coordination) — so make a "neutral VDA5050 gateway + marsupial-specific coordinated scheduling" the main thrust, with SEER RDS / Open-RMF as reference bases (see Coordination-software landscape).
• The data/eval directions (🟢2, 3) can be a second curve of "software capability extension," starting from the warehousing/haulage scenarios we already know.