Defensible

Defensible refers to the ability of a decision, action, or system to be justified, challenged, audited, and publicly explained through sufficient evidence and reasoning. A defensible system is not merely effective or safe: it must also support accountability, contestability, and governance.

Embodied

Embodied means that something or someone has a physical form in the real world, as opposed to being purely digital or virtual. Embodied agents can interact with their environment, other agents and humans in a tangible way.

Agents

Agent comes from “agere”, i.e., to act. An agent is an entity that can perceive its environment, make decisions, and take actions to achieve specific goals or objectives. It is autonomous, (possibly) self-sufficient, proactive, and (possibly) adaptive.

Modern embodied agents will increasingly operate in shared human environments:

• Schools
• Hospitals
• Public transportation
• Homes
• Workplaces

In these contexts, pure optimization is not enough.

An agent should not only ask:

• “What can I do?”

but also:

• “What should I do?”
• “What am I allowed to do?”
• “How should I behave with humans?”

This requires explicit representations of:

• Permissions
• Obligations
• Prohibitions
• Preferences
• Social conventions
• Ethical principles

Each stakeholder is modeled as a normative system:

• obligations
• permissions
• prohibitions
• contextual preferences

These normative systems are then used as sources of arguments.

Instead of relying on majority voting or fixed priorities, Jiminy tries to reach a moral agreement among stakeholders.

The framework resolves dilemmas through:

• interactions between arguments
• attacks and defenses
• argument acceptability semantics

Children are increasingly exposed to:

• conversational AI
• tutoring systems
• social robots

However, most existing AI systems are designed for adults.

This is particularly problematic in sensitive domains such as:

• education
• therapy
• special needs support

where safety, trust, and controllability become essential requirements.

The project develops a norm-first agent architecture where:

• LLMs provide flexible dialogue capabilities
• Norms constrain the reasoning cycle
• Machine-readable policies enforce safety checks

This enables behavior that is:

• Explainable
• Auditable
• Regulation-aware
• Compliant-by-design

Research challenges:

• Age-appropriate dialogue
• Safe interaction
• Controllability
• Robustness against unsafe outputs

Digital agents

• Mainly operate in virtual environments
• Interact through text or software APIs
• Can use software tools
• No/Limited physical awareness
• No direct interaction with the world

Embodied agents

• Perceive the physical world directly
• Interact through sensors and actuators
• Learn from situated interaction
• Affect real environments and humans
• Must reason about safety and uncertainty

This introduces new challenges

• Real-time decision-making
• Uncertainty and partial observability
• Safety-critical behavior
• Social interaction
• Accountability

Lifecycle

• offline
• online

Learning does not stop after deployment.

Embodied agents must continuously:

• adapt
• update beliefs
• refine behaviors

Tempo

• fast
• slow

Different decisions require different reasoning speeds.

Locus

• individual
• collective

Future AI will increasingly involve:

• coordination
• negotiation
• collective intelligence

Substrate

• subsymbolic
• symbolic

Neither paradigm alone appears sufficient.

Stage

• pre-action
• in-action
• post-action

Reasoning may happen:

• before acting
• while acting
• after acting