Safety-Critical Robot Data: Training Data for ISO 10218 and TS 15066 Compliance

Robots operating near humans must detect, predict, and react to human proximity with the reliability that safety standards demand. ISO 10218 and ISO/TS 15066 define the requirements; training data determines whether a learned perception system can meet them. The consequences of insufficient safety data are not degraded performance but physical harm.

Why Is Safety-Critical Robot Data Different from Standard Training Data?

Standard robot training data optimizes for task completion: pick success rate, navigation efficiency, manipulation accuracy. Safety-critical data must optimize for a fundamentally different objective: ensuring that the robot never causes harm to nearby humans, even at the cost of task performance. ISO 10218 defines safety requirements for industrial robot systems including protective stop, speed limitation, and safety-rated monitored stop functions. ISO/TS 15066 extends these requirements to collaborative robot applications where humans and robots share workspace, specifying permissible force and pressure limits for different body regions. Training a learned perception system to enforce these standards requires data that comprehensively covers the scenarios where safety interventions are necessary: humans entering the robot workspace from unexpected directions, partial occlusion behind equipment, unusual postures, and edge cases where human detection is most difficult and most critical.

[1][2]

What Makes Human Proximity Detection Data So Hard to Collect?

Collecting safety-relevant human proximity data presents a chicken-and-egg problem: you need data of humans near operating robots to train safety systems, but operating robots near humans without adequate safety systems is itself dangerous. SafetyBench evaluated the safety performance of robot policies across diverse scenarios and found that most failure modes involve edge cases where humans appear in unexpected positions relative to the robot workspace. DROID and Open X-Embodiment contain some instances of humans in the frame but without safety-specific annotations: body region proximity labels, approach velocity vectors, or occlusion state indicators. The data that safety systems need — systematic coverage of all directions of approach, partial visibility conditions, and rapid entry scenarios — is precisely the data that is most dangerous to collect without an already-functioning safety system.

[3][4]

How Do Current Datasets Fail Safety Certification Requirements?

Safety certification under ISO 10218 and TS 15066 requires demonstrating that the perception system can reliably detect humans across a defined range of conditions with a specified false negative rate. No existing public robot dataset was designed to support this certification process. The datasets that include human presence data — such as EPIC-KITCHENS for activity recognition or COCO for person detection — were not collected with safety-critical annotation requirements: body region segmentation at the resolution TS 15066 requires, approach velocity measurement, or systematic coverage of edge-case viewing conditions. Building a safety-certified learned perception system requires purpose-built data collection programs that systematically enumerate and capture the detection scenarios that certification testing will evaluate.

[1][2]

How Do Existing Datasets Support Safety-Critical Robot Training?

The table below compares data sources relevant to safety-critical robot perception against Claru custom collection. No open dataset was designed for safety certification; all require significant additional annotation and scenario coverage.

COCO (Person Detection)

Scale200K+ labeled images, 250K+ person instances
TasksPerson detection and pose estimation
EnvironmentsWeb-scraped images; diverse but uncontrolled
LimitationsNot robot-centric viewpoints; no proximity or approach velocity labels; no body region force-limit annotations; no systematic edge-case coverage

DROID

Scale76K trajectories with some human presence
TasksManipulation with incidental human observations
EnvironmentsResearch labs; humans occasionally in frame
LimitationsNo safety-specific annotations; human presence is incidental not systematic; no proximity zones or approach vectors

Open X-Embodiment

Scale1M+ trajectories; humans occasionally present
TasksManipulation tasks with bystanders in some datasets
EnvironmentsResearch labs across 22 platforms
LimitationsNo safety annotations; human presence is uncontrolled; no body region labels or force-limit mapping

SafetyBench (Sim)

ScaleBenchmark scenarios for robot safety evaluation
TasksSafety performance evaluation across diverse scenarios
EnvironmentsSimulated environments with controlled human models
LimitationsEvaluation benchmark only; simulation; no real-world training data; limited to scenario definitions

Claru Custom

Scale386K+ video clips, ~500 contributors, configurable safety-scenario protocols
TasksConfigurable: human proximity detection, approach direction coverage, body region annotation, occlusion edge cases, rapid entry scenarios
EnvironmentsReal workplaces, manufacturing floors, warehouses; actual human-robot shared workspace conditions
LimitationsRequires engagement lead time (days to launch, 1-2 week calibration); not a public benchmark; safety scenarios require careful protocol design
0+

Annotators

0

Countries

0M+

Annotations Delivered

Same-day

QA Turnaround

Frequently Asked Questions

Claru's safety data collection programs are designed to support certification under ISO 10218 (safety requirements for industrial robots) and ISO/TS 15066 (collaborative robot safety including force and pressure limits for 29 body regions). Data collection protocols enumerate the specific detection scenarios that certification testing evaluates and systematically capture coverage across the full scenario matrix.

Safety scenario coverage is tracked against a defined matrix covering approach directions, speeds, occlusion conditions, human postures, body types, and multi-person configurations. The collection protocol tracks completion percentage against each scenario dimension and identifies coverage gaps before data delivery. Edge cases where human detection is most difficult (partial occlusion, unusual postures, rapid entry) receive dedicated collection focus.

Yes. The annotation pipeline provides body region segmentation at the resolution ISO/TS 15066 requires, identifying the 29 specified body regions with per-frame labels indicating which regions are within the robot's proximity zones. This enables force-limiting controllers to set appropriate thresholds based on the specific body part at risk of contact.

Yes. Claru deploys data collection directly into target deployment environments. Safety risks are environment-specific, so training data must reflect the actual conditions: lighting, equipment layout, worker behavior, protective equipment, and proximity scenarios specific to the deployment site. Previous workplace programs have captured data across 10+ workplace categories in multiple countries.

// INITIATE

Your next hire isn't a vendor. It's a data team.

Tell us what you're training. We'll scope the dataset.

claru@contact ~ READY
CONNECTED
> Initialize consultation request...

Or email us directly at [email protected]

</>

References

  1. [1]International Organization for Standardization. “ISO 10218-1:2011 — Robots and Robotic Devices: Safety Requirements for Industrial Robots.” ISO Standard, 2011. Defines safety requirements for industrial robot systems including protective stop, speed limitation, and safety-rated monitored stop functions for human-robot coexistence. Link
  2. [2]International Organization for Standardization. “ISO/TS 15066:2016 — Robots and Robotic Devices: Collaborative Robots.” ISO Technical Specification, 2016. Specifies safety requirements for collaborative robot applications including permissible force and pressure limits for 29 body regions during human-robot contact. Link
  3. [3]Tung et al.. “SafetyBench: Evaluating the Safety of Robot Manipulation Policies.” arXiv 2024, 2024. Evaluated safety performance across diverse robot scenarios; found that most safety failures involve edge cases where humans appear in unexpected positions relative to the robot workspace. Link
  4. [4]Khazatsky et al.. “DROID: A Large-Scale In-the-Wild Robot Manipulation Dataset.” arXiv 2024, 2024. 76,000 manipulation trajectories with incidental human presence in some frames; no safety-specific annotations or systematic human proximity coverage. Link