Halim Djerroud

A Foundational Question

At the heart of my work lies a foundational question that goes far beyond classical problems of algorithmic optimization or local performance: how can we design robotic systems capable of making relevant, adaptive, and explainable decisions in real, dynamic environments that are inherently shared with humans?

This question engages fundamental reflection on the very nature of embodied artificial intelligence. Acting autonomously in the physical world does not merely consist of reacting to stimuli or maximizing an abstract reward function; it requires a structured understanding of the environment, the ability to anticipate its possible evolutions, and the capacity to inscribe the robot's action within a social, functional, and ethical framework comprehensible to humans.

Beyond Dominant Paradigms

This perspective has gradually led me to move away from two dominant paradigms of contemporary robotics, whose limitations become apparent when seeking to go beyond controlled or artificially simplified scenarios.

On one side, purely reactive architectures, stemming from behavior-based robotics and subsumption architectures, offer undeniable robustness and execution speed, but at the cost of an almost total absence of prospective capacity. A reactive robot can avoid an immediate obstacle, but is fundamentally incapable of reasoning about the delayed consequences of its actions, for example when an obstacle is mobile, manipulable, or socially mediated by a human.

On the other side, end-to-end deep learning approaches have demonstrated spectacular performance on perception or navigation tasks, but they rely on distributed and opaque representations, making the produced decisions difficult, if not impossible, to explain, justify, or audit. However, in contexts such as assistive robotics, human-robot collaboration, or shared environments, decisional opacity constitutes a major obstacle to acceptability, trust, and safety.

First Pillar: Multi-Level Representations

A first fundamental pillar of this vision lies in the use of multi-level representations, ranging from raw geometry to semantics and situated action. A robot cannot settle for a perception of the world reduced to pixels or unstructured point clouds.

• Geometric level: The world is described in terms of surfaces, volumes, obstacles, and navigable zones, anchored in explicit physical metrics.
• Topological level: These primitives are organized into structures capturing the connectivity of space independently of exact distances, allowing reasoning in terms of places, passages, and spatial relations.
• Semantic level: Perceived entities are categorized and connected by functional and social relations, opening the way to an interpretation of the world compatible with human language.
• Behavioral level: The representation encodes affordances and interaction constraints, that is, the possibilities for action offered by the environment according to context and implicit or explicit norms.

This stratification is not a simple stacking of independent layers, but the expression of a central cognitive hypothesis: abstraction is the key to generalization. A robot capable of reasoning about the abstract notion of controlled passage can transfer its skills between different types of doors without exhaustive relearning.

Second Pillar: Simulation as a Cognitive Mechanism

The second pillar of my approach concerns the central role of simulation in the decision-making process. In my conception, simulation is not a posteriori validation tool, but an a priori cognitive mechanism, directly involved in decision-making.

This approach is inspired by work on mental simulation and internal models in cognitive psychology and neuroscience, according to which biological agents evaluate the possible consequences of their actions by mentally exploring several plausible futures before acting.

Concretely, this idea is translated in my work through the use of multi-agent systems as internal simulation support. Entities perceived in the real environment are modeled as agents endowed with physical properties, action capacities, and specific constraints. The robot itself is integrated into this simulated world as a cognitive agent.

Before any engaging action, several scenarios are simulated, evaluated, and compared according to explicitly modeled criteria such as energy cost, collision risk, execution duration, or social acceptability. The actual execution of the action is then supervised by continuous comparison between predictions from simulation and actual observations.

Third Pillar: Cognitive Frugality and Meta-Decision

The third pillar, which today constitutes the most structuring axis of my research, is that of cognitive frugality and meta-decision. A truly intelligent system does not merely decide what to do; it must also decide how to decide.

Faced with a given situation, several decision-making strategies are possible, ranging from complete planning from symbolic models, to reusing past experiences, to adopting simple reactive behaviors, or even explicitly delegating the decision to a human. The choice between these strategies should not be arbitrary, but guided by a rational evaluation of criteria such as:

• Temporal urgency
• Task criticality
• Available computational and energy resources
• Human cognitive availability

This approach is part of the tradition of bounded rationality, as formulated by Herbert Simon and formalized by Russell and Wefald, while distinguishing itself through concrete anchoring in embedded robotics.

This cognitive frugality goes beyond simple technical optimization; it carries an ethical and ecological dimension. At a time when autonomous systems are multiplying, integrating computational parsimony as a design principle contributes to more sustainable and socially responsible robotics.

Methodological Choices

My methodological choices stem directly from this vision:

• Symbolic formalisms (PDDL): Allow explicit representation of preconditions, effects, and action constraints, making each decision traceable and justifiable.
• Multi-agent systems: Offer a natural framework for modeling environments populated by heterogeneous entities and supporting distributed simulation.
• RGB-D fusion: Constitutes a relevant compromise between perceptual richness and material constraints, allowing faithful apprehension of 3D geometry without resorting to heavy infrastructures.
• Transfer learning: Responds to realistic constraints of assistive robotics and industrial environments, where massive and exhaustive datasets are rarely available.