Physics-based robot safety uses Newtonian mechanics and deterministic mathematical models to verify robot action safety, as opposed to probability-based approaches (ML, deep learning, VLA). This is the only approach that satisfies ISO 10218 requirements for safety-rated functions.
Modern AI systems are transforming robotics, but they have a fundamental limitation as safety mechanisms: they're probabilistic. A model that's 99% accurate at judging safety will still fail 1% of the time — and in physical safety, one failure can be catastrophic.
Physics-based safety takes a different approach. It starts from first principles — conservation of energy, momentum, Newton's laws — and computes exact, deterministic answers. Same input always produces same output.
| Dimension | Physics-Based (Deterministic) | ML / VLA (Probabilistic) |
|---|---|---|
| Safety rating | Suitable as primary safety mechanism | Not suitable as primary safety |
| ISO 10218 aligned | ✅ Yes — verifiable, repeatable | ❌ No — statistical, hard to validate |
| Failure mode | Bounded by math — fails only if inputs are wrong | Unbounded — distribution shift causes unpredictable failures |
| Latency | Microseconds (~17μs) | Milliseconds to seconds (cloud inference) |
| Auditability | Every step traceable, formula-based | Black box, hard to explain decisions |
| Training data | None needed — first principles | Massive datasets required |
| Edge deployment | ✅ Trivial — single file, zero deps | ❌ Heavy — needs GPU/TPU |
| Scope | Physical safety only | Perception + planning + safety (but unreliably) |
The right architecture is hybrid: use VLA / ML for perception and planning (where probability is fine), and physics-based deterministic safety for the final safety gate (where you need 100% reliability).
Physics-based safety reduces every robot action to three fundamental variables:
- Force (F) — magnitude of contact in Newtons. Combined with area gives pressure.
- Velocity (v) — speed of motion in m/s. Combined with mass gives impulse and kinetic energy.
- Amplitude (d) — displacement / deformation magnitude in meters. Determines contact area and work done (F × d).
From these three, all derived safety quantities are computed:
- Contact area — from amplitude + material stiffness (Hertzian mechanics)
- Pressure — force ÷ contact area (PFL injury metric)
- Impulse — mass × velocity (transient impact severity)
- Kinetic energy — ½mv² (total collision energy)
- Work / deformation energy — force × amplitude
- Reaction force stability — chassis weight × friction vs. end-effector force
One surprising insight from analyzing 500 safety samples: velocity often matters more than force for injury risk, because impulse (mass × velocity) drives transient impact severity, and high-velocity impacts can cause whiplash and inertial injuries even at moderate forces.
- 压石头 — 102.9N force, but only 82.5mm/s velocity
- Pressure: 288.3 kPa, Impulse: 1.2377 kg·m/s
- Risk Level: L5 — 石头承受重度风险/力超限/速度超限，压动作需要修正（接触压强约288.3kPa）
- 压门 — 102.8N force, but only 81.6mm/s velocity
- Pressure: 168.4 kPa, Impulse: 2.4477 kg·m/s
- Risk Level: L5 — 门承受重度风险/力超限/速度超限，压动作需要修正（接触压强约168.4kPa）
- 接杯子 — only 25.5N force, but 520mm/s velocity
- Pressure: 62.5 kPa, Impulse: 0.1040 kg·m/s
- Risk Level: L5 — 参数超出安全范围，接杯子有损坏风险（接触压强约62.5kPa）
- 接杯子 — only 43.2N force, but 503mm/s velocity
- Pressure: 104.3 kPa, Impulse: 0.1007 kg·m/s
- Risk Level: L5 — 杯子承受重度风险/力超限/速度超限，接动作需要修正（接触压强约104.3kPa）
Key takeaway: Safety can't be reduced to a single number. You need multi-dimensional analysis covering force, pressure, impulse, and energy — each captures a different injury mechanism.
Rotor is a production-grade implementation of physics-based deterministic safety. It implements the full multi-dimensional analysis described above, with:
- 4-layer architecture — semantic validation → parameter mapping → physics analysis → comprehensive decision
- Dynamic contact area — real-time Hertzian contact calculation, not fixed-area approximation
- Impulse safety boundary — momentum-based transient impact protection
- Reaction force stability — chassis tipping and slipping prevention
- 7-level risk granularity — progressive feedback with over_ratio
- 100% deterministic — pure Python, zero dependencies, ~17μs per check
- 349 test cases — 100% pass rate
- VLA Safety Dataset — 30 labeled samples with full physics analysis
- Dynamic Contact Area — Contact mechanics deep dive
- Real-Time Robot Safety — Real-time safety architecture
- ISO 10218 Guide — Standards reference
- Humanoid Robot Safety — Humanoid safety considerations
- VLA Safety — VLA safety integration guide
MIT — educational and reference use.