Simin Luan

Trust in autonomous agents can shift from probabilistic alignment to a verifiable check at boot, ensuring they remain within their authorized limits regardless of learning.

Caste reassignment in robot swarms can now be auditable and externally governed, preventing privilege escalation through a novel asymmetric-trust protocol.

Consolidating knowledge in intelligent agents without altering their certified identity could revolutionize compliance in autonomous systems.

Decoupling capability names from versions lets you continuously deploy new robot skills without re-certifying the robot's core identity.

Replacing a single skill in a robot's compositional policy can swing task success by up to 50%, and existing behavioral distance metrics can't predict which skill matters most.

Current embodied AI benchmarks overlook critical governance aspects like controllability and safety during upgrades, a gap EmbodiedGovBench directly addresses.

Multi-robot coordination doesn't need to turn each robot into a fragmented multi-agent system; treating them as federated single agents actually improves governance and recovery.

Naive upgrades of embodied agent capabilities lead to unsafe activations in 60% of cases, but a governed upgrade framework can maintain task success while ensuring zero unsafe activations.

Guaranteeing safe and reliable behavior in embodied agents requires external runtime governance, not just smarter agents.

Forget retraining your robot from scratch: this capability-centric approach lets embodied agents learn new tricks without losing their core identity or safety guarantees.

Robots can achieve 100% task success and maintain safety across diverse tasks by modeling them as single agents with hot-swappable capability modules governed by a policy-separated runtime.