AMR Safety Simulation

This series has examined the components of navigation from occupancy grids through reactive avoidance. All of those components ultimately exist to answer one question — "Is it safe to deploy this AMR in a shared space with people?" AMR safety simulation answers that question in numbers.

The Point Where Concepts Become KPIs

When the navigation stack runs correctly, it becomes possible to measure the safety indicators that standards require.

Rendering diagram…

ISO 3691-4 — the safety standard for driverless industrial trucks and AMRs — requires collision avoidance, emergency stopping, and deceleration distance control. In simulation, these translate to:

• collision = 0 — zero collisions across all runs
• TTC (Time-To-Collision) > 2.0 s — always at least 2 seconds of clearance before any potential collision
• min_obstacle_distance > 0.5 m — minimum 0.5 m clearance from obstacles maintained at all times
• near_miss = 0 — no near-miss events

💡 "Moving ≠ passing the gate" — a robot reaching its destination does not mean it is safe. All four KPIs must simultaneously exceed their thresholds to pass the ISO 3691-4 gate. That fail-loud gate is what separates "it seemed to work fine" from "it is validated."

Not Once — Hundreds of Times. Statistics Are the Evidence

Safety cannot be guaranteed from a single run. The same scenario must be repeated hundreds to thousands of times with varied conditions (domain randomization) to produce statistical evidence.

Rendering diagram…

⚠️ Fleet makes it harder — when multiple robots operate in the same space, inter-robot collisions must also be zero. You need to cross-validate that a multi-AMR fleet maintains collision = 0 through round-robin yield behavior. Single-robot safety and fleet safety are different problems.

▶ Simulation footage — two Huskys placed head-on at either end of a 1.10 m gap. Passing each other needs 1.97 m, so only one can go through; the log records "HO_A: width 1.10 m < 1.97 m required to pass, HO_B has priority -> stop before entry", and HO_A then creeps up at 0.25 m/s as shown. Scored result for this run — collision 0, deadlock 0, minimum separation 0.932 m clear two of the gates, but minimum TTC of 0.23 s (threshold 2.0 s) and 2 near-misses (threshold 0) do not. The robots reached their goals and still failed the ISO 3691-4 gate — that is exactly where "working" and "validated" part company.

Closing the Series

From occupancy grid to here — modeling the environment as a grid, stacking a costmap, planning a path, tracking it, avoiding obstacles, and finally validating all of that against standard KPIs — this series has completed one full loop. Each navigation component ultimately converges on a single deliverable: a safety-validated AMR.

In a nutshell

📌 The navigation stack is ultimately validated against ISO 3691-4 safety KPIs (collision 0, TTC > 2 s, min_dist > 0.5 m, near_miss 0). "Working" is not "validated" — all KPIs must pass. Evidence requires not a single run but statistical results from hundreds to thousands of domain-randomized runs; for fleets, inter-robot collision must also be zero.