VLA safety is the missing layer between AI-generated actions and physical execution. Vision-Language-Action models excel at semantic understanding but lack physical intuition — they can't reliably estimate contact pressure, impulse, or reaction forces in real time.
Rotor provides a deterministic physics safety layer that sits between VLA output and robot actuators, ensuring every action stays within Newtonian safety bounds.
VLA models like Google Gemini Robotics ER 2 generate actions frame by frame based on visual input. They're trained on massive datasets, but they have fundamental limitations as safety-critical systems:
- Probabilistic output — 99% accuracy means 1% of actions may be unsafe. In the physical world, 1% failure rate is unacceptable for safety functions.
- Distribution gap — training data never covers every possible physical scenario. Out-of-distribution inputs produce unpredictable outputs.
- No physical intuition — VLA models "see" objects but don't intuitively understand force, pressure, impulse, or friction at the level of Newtonian mechanics.
- Latency mismatch — cloud inference (500ms+) is too slow for real-time safety, where you need sub-millisecond response.
- Auditability — when a VLA makes an unsafe decision, you can't trace why. ISO 10218 requires safety functions to be verifiable and explainable.
VLA Model → Action Plan → [Rotor Safety Layer] → Controller → Robot
↓
over_ratio feedback
↓
VLA replanning (progressive adjustment)
The VLA handles perception, planning, and semantics. Rotor handles the final physical safety gate. This follows the same architectural principle as industrial robots under ISO 10218: the safety system is separate from the control system, and it has veto authority.
From a dataset of 167 VLA-relevant robot action samples (humanoid, service, collaborative) across 3 scenarios (kitchen, living room, workshop):
| Risk Level | Count | Percentage | Meaning |
|---|---|---|---|
| L0 | 37 | 14.8% | 完全安全，裕度充足 |
| L1 | 38 | 15.2% | 很低风险，正常操作 |
| L2 | 31 | 12.4% | 低风险，消耗部分裕度 |
| L3 | 48 | 19.2% | 中风险，接近限值 |
| L4 | 35 | 14.0% | 中高风险，超出舒适区 |
| L5 | 33 | 13.2% | 高风险，接近超限 |
| L6 | 28 | 11.2% | 危险，必须拦截 |
Key insight: 41.3% of VLA-generated actions fall in L3 or above (medium to dangerous risk). A VLA model alone cannot be trusted to output safe actions — it needs a dedicated safety layer.
Humanoid robots introduce unique safety challenges: whole-body contact, balance stability, and operation in human environments.
举抽屉 — Risk Level L5
- Input: 10.5N force, 262.6mm/s velocity, 38.4mm amplitude
- Physics: 20.8 kPa pressure, 502 mm² contact area, 0.5252 kg·m/s impulse
- Why: 参数超出安全范围，举抽屉有损坏风险（接触压强约20.8kPa）
- Safe params: reduce force to 8.5N, velocity to 223.2mm/s
捏抱枕 — Risk Level L5
- Input: 7.5N force, 52.9mm/s velocity, 69.4mm amplitude
- Physics: 9.1 kPa pressure, 825 mm² contact area, 0.0264 kg·m/s impulse
- Why: 抱枕承受重度风险/速度超限，捏动作需要修正（接触压强约9.1kPa）
- Safe params: reduce force to 6.4N, velocity to 42.5mm/s
顶抽屉 — Risk Level L6
- Input: 154.4N force, 148.6mm/s velocity, 65.4mm amplitude
- Physics: 290.8 kPa pressure, 531 mm² contact area, 0.2972 kg·m/s impulse
- Why: 严重违规操作，顶抽屉必须立即停止（接触压强约290.8kPa）
- Safe params: reduce force to 85.0N, velocity to 85.0mm/s
| Feature | Value | Relevance to VLA |
|---|---|---|
| Latency | ~17μs per check | Fits within any control loop, no slowdown |
| Determinism | 100% | Safety is predictable, auditable, ISO 10218 aligned |
| Dependencies | Zero | Drop directly into VLA pipeline |
| Risk granularity | 7 levels | Progressive feedback, not just stop/go |
| Input modes | Natural language + JSON | Works with VLA text output or structured params |
| Semantic filter | Built-in | Catches impossible verb-object combinations |
| File size | ~140KB | Deploy anywhere, even on edge devices |
- VLA Safety Dataset — 30 labeled samples with full decision chain
- ISO 10218 Guide — Standards reference for robot safety
- Humanoid Robot Safety — Humanoid-specific safety deep dive
- Real-Time Robot Safety — Real-time safety architecture
- Dynamic Contact Area — Contact mechanics deep dive
- Physics-Based Safety — Why determinism matters for safety
- Robot Collision Detection — Collision detection methods
This project is provided for research and educational purposes only. It is NOT a certified safety device, NOT a substitute for professional risk assessment, and NOT intended for use in safety-critical applications without independent validation.
The authors make NO representations or warranties of any kind, express or implied, about the completeness, accuracy, reliability, or suitability of the software. Any use is at your own risk.
Always comply with local safety regulations (ISO 10218, ISO/TS 15066, and applicable national standards) and conduct thorough risk assessments before deploying any robotic system.
Underlying all VLA safety discussions is a simple physical truth: every action, no matter how complex the intent, resolves to three variables at the point of contact:
| Dimension | What It Measures | Why It Matters for Safety |
|---|---|---|
| Force | Contact pressure distribution | Directly determines tissue stress and injury threshold |
| Speed | Relative velocity at contact | Determines impulse transfer and reaction window |
| Amplitude | Movement range & displacement | Defines hazard zone size and contact geometry |
VLA models generate actions in high-dimensional semantic space. Safety happens in three-dimensional physical space. The Rotor Safety Engine bridges this gap by projecting every proposed action onto these three variables and computing the physical safety boundary in real time (~17μs per evaluation).
The safety layer is validated against 500+ labeled scenarios covering seven risk levels, but critically: it is not trained on them. The physics engine produces deterministic results from first principles, and the dataset serves as verification — not training material.
This is what makes it reliable: the same three variables that explain a kitchen collision also explain a workshop pinch and a humanoid fall. Physics is universal.
MIT — educational and reference use.