CEIAI Governance Database
Back to the database
Record STMT-W30EE95Voluntary

EmbodiedGovBench: A Benchmark for Governance, Recovery, and Upgrade Safety in Embodied Agent Systems

Report Institutional policyGlobalin force

EmbodiedGovBench is a governance-oriented benchmark for evaluating embodied agent systems across seven operational dimensions, moving beyond simple task success metrics. It tests unauthorized capability invocation, runtime drift robustness, recovery success, policy portability, upgrade safety, human override responsiveness, and audit completeness in both single-robot and fleet settings. The benchmark utilizes scenario templates, perturbation operators, and metric families to determine whether an embodied system remains a controllable, policy-bounded, and accountable runtime substrate.

Summary and content fields are drawn from the document text and are pending editorial review.

Who issued it

Harbin Institute of TechnologyAcademic · CN
Lead issuer
Heriot-Watt UniversityAcademic · MY
Co-issuer
Soochow UniversityAcademic · CN
Co-issuer
arXivAcademic
Publisher
Legal forceVoluntary
Under whose authorityInstitutional policy
Judged fromfull text

Who it governs

Researchers and academicsRegulatorsAI developers and providersAI deployers and users

AI researchers, roboticists, and evaluation engineers looking to measure and ensure the governability, safety, and accountability of embodied agent systems.

Sectors
Research and science · embodied AI and robotics researchCross-sector

What the document contains

What's covered

The document covers a complete benchmark framework for evaluating the operational governance of embodied agent systems, including formal definitions of seven governance dimensions, metrics, protocol families, baseline evaluation matrices, harness architectures, and empirical pilot experiment results.

What's missing

It does not provide real-world hardware deployment validations outside of simulation environments, nor does it establish binding legal regulations or standards compliance certifications.

Key takeaways
  • Traditional embodied AI benchmarks measure task success but leave capability governance unmeasured.
  • EmbodiedGovBench introduces seven governance dimensions spanning legality, runtime drift, recovery, portability, upgrades, human override, and audit completeness.
  • The benchmark includes both single-robot and fleet tracks with parameterized scenario templates and perturbation operators.
  • Pilot experiments show that task success can mask severe governance failures, and governance-aware systems incur a performance trade-off for policy compliance.
  • A governance-aware scoring engine produces multi-family scores (Gcap, Grec, Gevo, Gacct) combined into an overall GovScore.
Principles it states
The field should evaluate embodied systems not only by what they can do, but also by whether their capabilities remain governable under execution, failure, and evolution.Governance is multi-dimensional and requires stress-oriented, trace-level evaluation rather than outcome-level scoring alone.Absence of governance failure is not governance success, and negative governance events should be penalized.
Outline
· 1 Introduction
· 2 Related Work
· 3 Governance Benchmark Dimensions
· 4 Benchmark Structure and Protocols
· 5 Metrics and Scoring
· 6 Baselines and Evaluation Protocol
· 7 Prototype Benchmark Instantiation
· 8 Experiment
· 9 Discussion
· 10 Conclusion
Issued13 April 2026
Length14,637 words · 34 pages
FormatNot recorded
Scopeglobal

How it scores

Pending review

Each score is backed by a quote verified against the document text. Editorial review has not yet confirmed them.

Fairness

0/5

Non-discrimination and equitable treatment

screening score · 2026-10-23

Human Dignity

0/5

Respect for the inherent worth of persons

screening score · 2026-10-23

Privacy

0/5

Protection of personal data and private life

screening score · 2026-10-23

Transparency

4/5

Openness about how AI systems work and are used

After execution, failure, or escalation, can the system reconstruct a sufficient operational trace to support attribution, debugging, governance review, and accountability

3.8 Audit Completeness

screening score · 2026-10-23

Enforcement

4/5

Consequences and compliance mechanisms with real force

Each governance dimension is measured by a small set of metrics computed per scenario instance

5.1 Individual Dimension Metrics

screening score · 2026-10-23

Topics

Physical AI Governancesupporting
embodied_ai runtime_governance benchmark robotics_safety auditability