An RIS is an array of many low‑power meta‑atoms that can impose programmable phase (and sometimes amplitude/polarization) shifts on impinging radio waves. By coordinating these shifts, the RIS shapes the wireless channel—e.g., redirecting energy toward desired users or nulling interference—without performing conventional RF amplification.

Active relays and small cells receive, process, and re‑transmit signals using power‑hungry RF chains. An RIS performs near‑passive reflection/redirect with orders‑of‑magnitude lower power, lower cost, and lower noise figure—but also with limited per‑element control (quantized phase, no baseband processing) and no intrinsic gain.

RIS adds controllable propagation paths to overcome blockage, fill coverage holes, and improve SINR with minimal energy footprint. In EXACT‑6G, RIS is a physical‑layer lever that our higher‑layer orchestration can schedule alongside compute and spectrum to meet KPIs (throughput, latency, reliability, energy efficiency).

Intuitively, we choose each element’s phase to align the reflected wavefront at the target, like turning many tiny mirrors. For a far‑field target, a common rule is to set element *n* to cancel the path‑length phase: ϕₙ ≈ −k·Δℓₙ (mod 2π), where k = 2π/λ and Δℓₙ is the extra distance via element *n*. Summing the coherent reflections yields a main lobe toward the desired angle and weaker sidelobes elsewhere.

Far‑field models assume plane waves and angle‑only steering; near‑field (spherical) models matter when users are within the surface’s Rayleigh distance, so focusing depends on both range and angle (“beamfocusing”). Near‑field control can create spatial hotspots for higher SNR or secrecy but needs more accurate geometry and phase design.

Options include: (i) cascaded channel estimation using pilot patterns with ON/OFF or coded RIS states; (ii) compressed sensing exploiting channel sparsity; (iii) geometry‑aided methods using position/orientation; and (iv) learning‑based predictors. Choice depends on mobility, RIS resolution, and overhead budgets.

A controller programs the RIS via low‑rate control links. EXACT‑6G targets control that respects 5G timing and signaling constraints (e.g., slot‑aligned updates, bounded configuration latency) so that RIS states can be coordinated with gNB scheduling, beam management, and edge compute actions.

Common constraints: quantized phase (e.g., 1–3 bits), coupled amplitude/phase, limited refresh rate, mutual coupling, and per‑tile grouping. These limit the ideal beam shape and update speed, so control/orchestration must plan around quantization and latency.

RIS consumes milliwatts to a few watts (controller + drivers), far below small cells. That enables greener coverage shaping and physical‑layer interference mitigation with a small carbon footprint—central to EXACT‑6G’s sustainability goals.

If mobility is very high, control latency/overhead can outweigh gains. Severe NLOS without viable geometric paths, strong hardware impairments, or tight secrecy/accuracy demands under coarse quantization can also limit benefits.

It’s the seamless fabric spanning cloud, edge (metro/regional), and far‑edge/on‑prem (e.g., gNB, RIS controller, or MEC nodes). Workloads (RAN functions, analytics, AI models) can be placed and migrated across this fabric based on latency, reliability, cost, and energy constraints.

Because KPIs vary over time (traffic, mobility, channel conditions). Orchestration decides “what runs where and when”—e.g., placing channel prediction near the far‑edge for low latency while training heavy models in the cloud—so that the end‑to‑end service meets SLA targets.

Decisions (e.g., scaling a microservice, switching an RIS configuration policy, re‑routing slices) are accompanied by human‑interpretable reasons or attributions (features, constraints, rules). This boosts trust, debuggability, and compliance, especially when AI/ML or MARL policies are involved.

• Far‑edge: fast loops—beam/RIS selection, link adaptation hints, lightweight anomaly detection.
• Edge: multi‑cell coordination, short‑horizon predictions, local digital twins.
• Cloud: global optimization, long‑horizon training, dataset curation, inter‑domain coordination.

By co‑optimizing placement (close to users for tight loops), redundancy (active/standby instances), and network paths (SRv6/TSN where applicable). Orchestration enforces budgets per service (e.g., inference deadlines) and moves workloads proactively when conditions shift.

Use interpretable policies (rules/constraints), post‑hoc attribution (e.g., feature importance for a scaling action), and policy introspection for MARL (e.g., counterfactuals explaining why a policy chose configuration A vs. B). Explanations are logged and surfaced via dashboards for operators and researchers.

Where possible, sensitive features remain at the far‑edge with privacy‑preserving aggregation (e.g., federated updates). Access control, audit trails, and anonymization protect datasets used for training and troubleshooting.

Open and interoperable abstractions are favored—e.g., O‑RAN for RAN control apps, ETSI MANO‑style lifecycles for VNFs/CNFs, and telemetry via open APIs. EXACT‑6G prototypes align with these ecosystems to ease adoption.

RIS is exposed as a controllable network function with states and KPIs. The orchestrator schedules RIS reconfigurations alongside RAN resource assignments and compute placement, ensuring that RIS updates are synchronized with beam management and do not violate timing/overhead constraints.

A demonstrable reduction in energy per bit and improvements in QoS/QoE via joint optimization of RIS, RAN, and compute placement— for key decisions so that operators can validate, audit, and trust the system.