SkillGate: every robot skill safety-verified before it moves on the factory floor
The problem
Vision-language-action models let robots learn new skills in weeks, but a skill that works in simulation can still do something unsafe around people — and from 20 January 2027 the EU Machinery Regulation requires third-party assessment for machine-learning safety functions. Manufacturers have no way to verify a downloaded or fine-tuned robot skill, and certification today is a manual engagement that takes weeks per robot cell.
The idea
Why now
Gemini Robotics 2.5 proved VLA skills deploy to live Mercedes production lines on 10 November, DeepMind open-sourced the ASIMOV-Jailbreak benchmark the same day, the EU Machinery Regulation deadline is ten weeks out, and TÜV SÜD with Accenture showed simulation-based safety checks at GTC Berlin in October — the market is already buying safety evidence for physical AI.
What it combines
Gemini Robotics 2.5's deployable robot skills plus ASIMOV-Jailbreak's open safety benchmark plus a Strands-Decider-class cheap triage model. The model alone ships skills, the benchmark alone tests prompts, and triage alone routes — together they make verification fast and cheap enough to run on every skill version, not once a year per cell.
MVP
Weekend MVP: a CLI that takes a task spec, generates 50 ASIMOV-style jailbreak prompts, runs them against a hosted VLA model in simulation, and emits a pass/fail report with the failing prompts attached. Skip: digital-twin integration, notified-body paperwork, and multi-robot support.
How it makes money
The moat
The verification corpus is a regulatory-data advantage: every skill tested grows a library of failure modes and jailbreak prompts that makes the next verification stronger — a dataset no incumbent holds for VLA skills. Robot OEMs will not certify third-party skills because it transfers liability to them, and hardware safety incumbents sell interlocks, not AI-skill verification — a different business model.
Distribution
B2B to robotics integrators, AMR and humanoid deployers, and manufacturers. The wedge is EU manufacturers facing the January deadline, reached through the integrators who deploy the skills and already own the customer relationship.
Why it wins
TÜV SÜD and Accenture's Physical AI Safety Suite verifies whole cells in simulation over weeks; SkillGate verifies the skill itself — the software artifact — in hours, per version, with a machine-readable report. Incumbent certifiers sell engagements; this sells a check.
Risks
Simulation-to-reality gap: a skill that passes in simulation can still fail on the floor. The MVP de-risks it by validating against one real robot cell's logged incidents before claiming coverage.
Build it with
- Gemini Robotics 2.5: DeepMind's most capable robot model, tested on a live Mercedes lineThe deployable VLA skills that need verifying — proven on a live Mercedes line.
- Strands Decider: AWS's open-weight decision model that picks instead of generatingCheap triage: which safety scenarios need full simulation and which get quick checks.
- Qwen4-Max Preview: Alibaba's next flagship, open for 14 days only1M context to ingest a skill's full safety file, task spec and incident history in one pass as the reviewing judge.
Repo to start from
skillgate: a CLI and service that safety-verifies vision-language-action robot skills against jailbreak benchmarks and simulation before factory deployment.
Evidence
- Gemini Robotics 2.5: Accelerating Physical Intelligence
- Machinery Regulation 2023/1230 vs Machinery Directive: what changes in 2027
- TÜV SÜD industrial robotics testing and certification
- How TÜV SÜD and Accenture are simulating robot safety checks
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.