Design Book › Safety and agent access

Safety and agent access

What a designer must do

  1. Keep hands out of pinch zones while motors are powered [1].
  2. Be ready to stop with CTRL+C or by unplugging power before you run a policy [2]. Its check times a stop from a running policy.
  3. Write a risk assessment that includes the tool and the payload [3], [4].
  4. Do not invent contact force limits. Get them from a licensed copy of the current standard [5].
  5. Write down what an agent may do to the robot [6]. Enforce it outside the agent, as the access-limits guideline asks [7].
  6. Keep a person within reach of a physical power cut during agent-driven tests [8].

Scope: the safety and agent-access guidelines. This chapter is a design aid, not a conformity statement. Most safety guidelines come from public summaries, not from the paid standards. The tables say which.

How much to trust each guideline

Every safety guideline based on an ISO document, vendor blog, vendor page or agency summary has level guide [9], [10]. That reflects what was read. The must safety guidelines come only from a regulation text [11] and a hardware course [1].

How much to trust each guideline
Source, as the guideline states itWhat was readSourcesStanding
ISO 10218-1:2025ISO store public summary. Standard text not read[9]Public summary of a paid standard
ISO 10218-2:2025ISO store public summary. Standard text not read[12]Public summary of a paid standard
ISO/TS 15066:2016ISO store public abstract. Specification text not read[13]Public abstract of a paid specification
ISO 13482:2014ISO store public abstract. Standard text not read[14], [15]Public abstract of a paid standard
ISO 13849-1:2023ISO store public abstract. Standard text not read[16], [17]Public abstract of a paid standard
ISO 9283:1998ISO store listing. Standard text not read[18]Title listing only
ISO/PAS 5672Named on an IFA page. Paid document not read[19]Pointer only
Robotiq vendor blogSecondary. Some rows predate ISO/TS 15066 and cite ISO 10218-2:2011; others are undated in the capture[20], [10], [3], [21], [22], [23], [24]Vendor summary, not a standard
IFA (DGUV) pagesSummary pages and a webinar extract (ISSA, 2024-06-24)[25], [26], [5], [19], [27], [4], [28]Agency summary
NIOSH robotics overview, updated 2026-10-28Overview page[29], [30], [31]Agency summary
NIST project pageSummary of program aims[32]Agency summary
Universal Robots marketing pageSecondary[33]Vendor claim
29 CFR 1910.147 and 1910.212, at Cornell LIIRegulation text as republished. US workplaces[11], [34]Regulation text; applies where it applies
NVIDIA SO-101 sim-to-real course, latest alias, no version shownCourse pages[1], [35], [36], [37], [2]Vendor course for this arm
Squint, arXiv:2602.21203Paper[38]One paper

Bench operation

Bench operation
LevelRequirementCheckSource
mustNo hand inside a zone where links can trap it while motors are poweredreview: brief operators before power-up[1]
mustLabel the SO-101 leader and follower supplies. They differ at 5 V and 12 V and must not be swappedmeasure: read each output with a meter[35]
shouldRoute cables so they cannot snag during motionmeasure: jog full range at low speed, watch them[36]
shouldInspect for damage and secure cables before each power-upreview: signed pre-power checklist[37]
mustOn an SO-101 bench setup, be ready to stop with CTRL+C or by unplugging power before running a policymeasure: time a stop from a running policy[2]
guideLimit motion on a first run of a sim-trained policy. Squint scales actions down for safety. From a papermeasure: log peak joint speed[38]

Risk assessment

All rows are guide. They come from a vendor blog or agency pages, not from a standard's text.

Risk assessment
RequirementSourceCheckSource
Complete an assessment for any collaborative robot before productionVendor blogreview: signed assessment[10]
Include grippers, end effectors and accessoriesVendor blogreview: lists each attached device[3]
Rate each hazard by harm, likelihood, ability to avoid it and exposure frequencyVendor blogreview: each row has four ratings[21]
Treat the tool and carried object as part of the collaborative robotIFA pagereview: covers tool and payload contact[4]
Assess the injury risk that remains after protective measuresIFA pagereview: residual risk per hazard[28]
List struck-by, caught-between, crushing, trapping, slip, trip, fall and electrical hazardsNIOSH pagereview: covers each class[29]
Robot contact is a real injury source. NIOSH counted 41 US robot-related fatalities in 1992-2017NIOSH pagereview: plan signed before powered tests near people[30]
Re-review as technology and application changeNIOSH pagereview: a change triggers a new review[31]
Power-and-force-limiting cobots usually still require an assessment, even when no additional safety devices are neededVendor blogreview: exists for the application[20]
A vendor's certified-safety claim covers what it names, such as the robot's safety system, and not your applicationVendor marketing pagereview: application has its own assessment[33]

Contact forces

This book holds no body-region force or pressure limit. The guidelines say to take them from a licensed copy of the current standard. A number written here would be invented [5].

Contact forces
RequirementSourceCheckSource
Check both total contact force and local peak pressureIFA pagemeasure: at each body region at risk[25]
Measure with a calibrated device, not by feel. Pain thresholds vary between people and over timeIFA pagemeasure: calibrated force and pressure device[26]
Take limits from a licensed copy. IFA says verified limits are moving into EN ISO 10218 Part 2IFA pagereview: each limit cites standard, edition, table[5]
IFA reports that ISO/PAS 5672, published at the end of 2023, sets out how to measure the force and pressure when a robot touches a person. Its text was not read hereIFA pagereview: report names its method[19]
In collaborative use, give tools and links no sharp edges or pointsIFA webinar extractreview: inspect tool edges and fingertips[27]

Stops and separation

All rows are from a vendor blog that predates the ISO/TS 15066 release. Treat them as vocabulary, not as requirements.

Stops and separation
RequirementCheckSource
A safety-monitored stop has brakes on. It is not a shutdownmeasure: robot stays energized and still[23]
Require a go-ahead before resumingmeasure: no motion until the go-ahead[24]
Under speed and separation monitoring, verify that the robot slows in each zone and stops when a person comes too closemeasure: walk a target through each zone[22]

The database holds no brake fact for the STS3215. Do not assume the safety-monitored stop above describes your arm [23].

Standards: scope only

These rows tell you what a standard covers, as its public summary says. They do not tell you what it requires.

Standards: scope only
StandardWhat the summary saysCheckSource
ISO 10218-1:2025Classes the industrial robot as partly completed machinery. A robot's own features are not the whole safety casereview: a separate safety file exists[9]
ISO 10218-2:2025Covers integration, commissioning, operation, maintenance and decommissioningreview: a section per phase[12]
ISO/TS 15066:2016Does not apply to non-industrial robots. Do not claim conformity for one; the principles may still inform designreview: claims match scope[13]
ISO 13482:2014Limited to earthbound robotsreview: robot class matches scope[14]
ISO 13482:2014Excludes robots faster than 20 km/h and robot toysmeasure: top speed against the limit[15]
ISO 13849-1:2023Does not specify an application's safety functions or required performance levels (PLr). Machine-specific requirements can be considered in a Type-C standardreview: each function cites its level and source[16]
ISO 13849-1:2023Gives no specific security measures, and notes that security issues can affect safety functionsreview: covers each remote command path[17]
ISO 9283:1998By its title, the standard sets performance criteria for industrial manipulators and the methods to test themreview: accuracy figures cite a method[18]
NIST projectShow safety with test methods and metrics, not assertionsmeasure: name the method behind each claim[32]

Law

These two rows apply where the regulation applies. They come from 29 CFR as republished by Cornell LII.

Law
LevelRequirementCheckSource
must29 CFR 1910.147: control hazardous energy before servicing a robot that could start unexpectedly or release stored energyreview: written energy-control procedure[11]
must29 CFR 1910.212: guard a machine's point of operation if its operation exposes an employee to injuryreview: guard or alternative is documented[34]

Agent access: where an agent may sit

An agent is a new command source. Treat its position as a boundary: name the boundary, and check it [6]. The guidelines on agent access, agent commands and agent control come from Anthropic's own write-ups, captured 2026-11-08. The database has no independent source for them. The two guidelines on reading docs come from the Newton 1.7.0 and Isaac Lab 3.0.0 docs.

Agent access: where an agent may sit
PositionWhat the guideline saysSourceLevel
Real-time balance loopDo not. Real-time legged control needs about 83 Hz; non-reasoning inference in that study ran at about 0.2-0.4 Hz[39]must
Driving joints directlyModels mostly fail at this[40]should
Supervising a pretrained controllerPreferred. Use the highest usable abstraction[40]should
Writing commands or code for hardwareCheck travel against clear distance, outside the agent[41]must
Reading docs for a robotVerify defaults and API names against the installed release[42]must

Agent access: limits

Agent access: limits
LevelRequirementCheckSource
mustAccess level is a core part of the safety review. Small tool changes can make large capability changesreview: written down, reviewed per change[6]
guideAnthropic says better ways are needed to grant physical access with clear limits, so a system can affect some objects while blocked from othersreview: allow-list of motions and objects, enforced outside the agent[7]
shouldA stationary arm under agent control is not low risk. Anthropic calls it a plausible deployment with direct safety relevancereview: same access review as mobile robots[43]
mustBefore a timed motion, compute speed times duration against clear distance. A Project Fetch team sent a robot dog at people through this mistakereview: travel check outside the agent[41]
mustKeep a person within reach of a physical power cut. Project Fetch's organizer had to grab the robot and power it offmeasure: rehearse and time it[8]
mustRead docs at the pinned simulator versionreview: plan names exact versions[44]
mustNewton's tuning guide tells humans and agents that the code is the source of truthreview: defaults cite a file or pinned doc[42]
shouldVerify connection facts against the hardware. Online information varies in accuracymeasure: link and sensor data on the real robot[45]

Agent access: test it

Agent access: test it
LevelRequirementCheckSource
mustA result from a paused simulator is an upper bound, not real-time performancereview: reports state whether it was paused[46]
shouldTest the supervising agent with deliberately wrong policy proposals. Models still degradesim: inject bad actions, score recovery[47]
shouldProbe whether the agent notices silently corrupted commandssim: inject drift, score detection[48]
shouldDo not rely on agent vision and memory alone for real navigation. All tested models failed one hallway loop in an informal testmeasure: supervised real test before untended use[49]
shouldWhen quoting results, state the control interface. The same model scores differently across interfacesreview: names interface, robot and task[50]
guideCode volume is not progress. The Claude team in Project Fetch wrote about 9 times more code; the authors call some of it arguably a distractionreview: track task metrics[51]
guideProject Fetch was one experiment with two teams. Do not generalize it to reliability claimsreview: claim states the sample size[52]

On the SO-101 with an agent

Where the database is thin

Sources

Each guideline in this chapter with its source, the date it was read and the exact words of the source. Level: must means breaking it causes failure or harm, should is good practice, guide is a rule of thumb.

  1. While motors are powered, no hand may be inside a zone where links can trap it. Level: must. docs.nvidia.com 2026-11-08
  2. On an SO-101 bench setup, be ready to stop with CTRL+C or by unplugging power before running a policy. Level: must. docs.nvidia.com 2026-11-08
  3. Include grippers, end effectors and accessories in the risk assessment. From a vendor blog summary, not a standard. Level: guide. blog.robotiq.com 2026-11-08
  4. Treat the tool and carried object as part of a collaborative robot in the assessment. From an IFA summary page. Level: guide. dguv.de 2026-11-08
  5. Do not invent body-region contact limits; take them from a licensed copy of the current standard, because IFA says the verified limits are moving into EN ISO 10218 Part 2. From an IFA summary page. Level: guide. dguv.de 2026-11-08
  6. Treat an agent's access level to a robot as a core part of the system's safety review, because small tool changes can make large capability changes. Level: must. anthropic.com 2026-11-08
  7. Anthropic says better ways are needed to grant physical access with clear limits, so a system can affect some objects while blocked from others. Level: guide. anthropic.com 2026-11-08
  8. Keep a person within reach of a physical power cut during agent-driven hardware tests; Project Fetch's organizer had to grab the robot and power it off. Level: must. anthropic.com 2026-11-08
  9. Do not treat a robot's own safety features as the whole safety case, because ISO classes the industrial robot as partly completed machinery. From the ISO store summary, not the standard text. Level: guide. iso.org 2026-11-08
  10. Complete a risk assessment for any collaborative robot before production. From a vendor blog summary, not a standard. Level: guide. blog.robotiq.com 2026-11-08
  11. Where 29 CFR 1910.147 applies, control hazardous energy before servicing a robot that could start unexpectedly or release stored energy. Level: must. law.cornell.edu 2026-11-08“the unexpected energization or start up of the machines or equipment, or release of stored energy could cause injury to employees.”
  12. Cover integration, commissioning, operation, maintenance and decommissioning in a robot cell's safety plan. From the ISO store summary of ISO 10218-2, not the standard text. Level: guide. iso.org 2026-11-08
  13. Do not claim ISO/TS 15066 conformity for a non-industrial robot; its principles may still inform design. From the ISO abstract, not the text. Level: guide. iso.org 2026-11-08
  14. Check ISO 13482 scope before you cite it: it is limited to earthbound robots. From the ISO abstract, not the text. Level: guide. iso.org 2026-11-08
  15. Note that ISO 13482 excludes robots faster than 20 km/h and robot toys. From the ISO abstract, not the text. 20 km/h Level: guide. iso.org 2026-11-08
  16. ISO 13849-1 does not specify an application's safety functions or PLr; machine-specific requirements can be considered in a Type-C standard. Level: guide. iso.org 2026-11-08
  17. ISO 13849-1 gives no specific security measures, and it notes that security issues can affect safety functions. Level: guide. iso.org 2026-11-08
  18. This guideline is held back: its source does not allow us to publish it yet.
  19. This guideline is held back: its source does not allow us to publish it yet.
  20. Power-and-force-limiting cobots usually still require a risk assessment even when no additional safety devices are needed. Level: guide. blog.robotiq.com 2026-11-08
  21. Rate each hazard by harm, likelihood, ability to avoid it and exposure frequency. From a vendor blog summary, not a standard. Level: guide. blog.robotiq.com 2026-11-08
  22. With speed and separation monitoring, verify that the robot slows in each zone and stops when a person comes too close. From a vendor blog summary, not a standard. Level: guide. blog.robotiq.com 2026-11-08
  23. Treat a safety-monitored stop as a stop with brakes on, not a shutdown. From a vendor blog summary, not a standard. Level: guide. blog.robotiq.com 2026-11-08
  24. Require a go-ahead before resuming from a safety-monitored stop. From a vendor blog summary, not a standard. Level: guide. blog.robotiq.com 2026-11-08
  25. Check both total contact force and local peak pressure for a person-robot collision. From an IFA summary page. Level: guide. dguv.de 2026-11-08
  26. Measure contact loads with a calibrated device, not by how a contact feels, because pain thresholds vary between people and over time. From an IFA summary page. Level: guide. dguv.de 2026-11-08
  27. This guideline is held back: its source does not allow us to publish it yet.
  28. Assess the injury risk that remains after protective measures, as a residual risk stays. From an IFA summary page. Level: guide. dguv.de 2026-11-08
  29. Include struck-by, caught-between, crushing, trapping, slip, trip, fall and electrical hazards in the hazard list. From a NIOSH overview page. Level: guide. cdc.gov 2026-11-08
  30. Treat robot contact as a real injury source, and not as a formality: NIOSH counted 41 US robot-related fatalities in 1992-2017. From a NIOSH overview page. 41 fatalities Level: guide. cdc.gov 2026-11-08
  31. Re-review robot risk as the technology and the application change. From a NIOSH overview page. Level: guide. cdc.gov 2026-11-08
  32. Show that collaborative systems are safe and correct with test methods and metrics, not with assertions. From a NIST project page, a summary of program aims. Level: guide. nist.gov 2026-11-08
  33. Read a vendor's certified-safety claim as covering what it names, such as the robot's safety system, and not your application. Secondary: a vendor marketing page. Level: guide. universal-robots.com 2026-11-08
  34. Where 29 CFR 1910.212 applies, guard a machine's point of operation if its operation exposes an employee to injury. Level: must. law.cornell.edu 2026-11-08“The point of operation of machines whose operation exposes an employee to injury, shall be guarded.”
  35. Label the SO-101 leader and follower supplies, because they differ at 5 V and 12 V and must not be swapped. 5 and 12 V Level: must. docs.nvidia.com 2026-11-08
  36. Route cables so that they cannot snag during motion. Level: should. docs.nvidia.com 2026-11-08
  37. Inspect the robot for damage and secure cables before each power-up. Level: should. docs.nvidia.com 2026-11-08
  38. Limit motion on a first hardware run of a sim-trained SO-101 policy; Squint scales actions down for safety. Level: guide. arxiv.org 2026-11-08
  39. Do not put a language model in a real-time balance loop: real-time legged control needs about 83 Hz, against about 0.2-0.4 Hz for non-reasoning inference in that study. 83 Hz Level: must. anthropic.com 2026-11-08
  40. Prefer an agent supervising a pretrained controller over an agent driving joints directly, because models mostly fail at direct joint drive. Level: should. anthropic.com 2026-11-08
  41. Before an agent or its code sends a timed motion, compute speed times duration against clear distance; a Project Fetch team sent a robot dog at people through this arithmetic mistake. Level: must. anthropic.com 2026-11-08
  42. Have agents verify default values and API names against the code of the installed release, as Newton's tuning guide tells humans and agents to do. Level: must. newton-physics.github.io 2026-11-08“the code is the source of truth.”
  43. Do not treat a stationary arm under agent control as low risk, because Anthropic calls it a plausible deployment with direct safety relevance. Level: should. anthropic.com 2026-11-08
  44. Have agents read docs at the pinned simulator version, because the Isaac Lab 3.0.0 project notes that compatibility with later releases has not been validated. Level: must. isaac-sim.github.io 2026-11-08“or later versions has not been validated.”
  45. Have an agent verify robot connection facts against the hardware, because online information varies in accuracy. Level: should. anthropic.com 2026-11-08
  46. Treat agent results from a paused simulator as an upper bound, not as real-time performance. Level: must. anthropic.com 2026-11-08
  47. Test the supervising agent with deliberately wrong policy proposals, because models still degrade when those policies are wrong. Level: should. anthropic.com 2026-11-08
  48. Probe whether an agent notices silently corrupted commands, as Anthropic's evaluation does, before trusting its closed-loop self-monitoring. Level: should. anthropic.com 2026-11-08
  49. Do not rely on agent vision and memory alone for real navigation: all tested models failed one hallway loop in Anthropic's informal test. Level: should. anthropic.com 2026-11-08
  50. When quoting agent robotics results, state the control interface, because the same model scores differently across interfaces. Level: should. anthropic.com 2026-11-08
  51. Do not read agent code volume as progress: the Claude team in Project Fetch wrote about 9 times more code, and the authors call some of it arguably a distraction. 9 times Level: guide. anthropic.com 2026-11-08
  52. Cite Project Fetch as one experiment with two teams, and do not generalize it to reliability claims. Level: guide. anthropic.com 2026-11-08
  53. Do not treat Isaac Lab joint_velocity_limit as a safety clamp, because its enforcement depends on the backend. Level: must. isaac-sim.github.io 2026-11-08“is not a backend-independent safety clamp”
  54. Feetech broadcast and sync writes normally return no status, so they cannot confirm action by one servo. Level: must. github.com 2026-11-08
  55. This guideline is held back: its source does not allow us to publish it yet.