AI for autonomous defense drones & surveillance

A three-day, hands-on workshop to build ethical, safe, auditable AI for non-weaponized autonomous drones in public-good missions—turning governance and privacy-by-design into requirements and hazard logs, making perception robust with calibration/OOD checks and model cards, and implementing human oversight, safe comms, geofencing, and coordinated multi-UAS behaviors—so you leave with concrete templates, test evidence, and operator SOPs.

Enable participants to design, test, and operate ethical, safe, and auditable AI for non-weaponized autonomous drones, translating policy into concrete technical controls and operator procedures.

• Turn governance, ethics, and privacy-by-design into actionable requirements and controls.

• Build a policy-aware requirements matrix, hazard log, and clear geofence/abort criteria.

• Design robust perception with calibration & OOD detection to manage false alarms/misses.

• Create evaluation plans and model/system cards that document limits and safe use.

• Implement privacy-preserving sensing and human-on-the-loop supervision.

• Engineer fail-safe comms, logging/traceability, and coordinated multi-UAS behaviors.

• Compile an assurance case with test evidence and operator SOPs for audits/field trials

📅 Day 1 – Governance, Ethics & Compliance for Autonomous Aerial Systems

• Legal & ethical foundations: meaningful human control, safety-of-life priorities, privacy/civil-liberties by design
• Policy alignment: risk assessments, impact assessments, auditability, documentation (model/system cards)
• Data governance: consent, provenance, bias/representativeness, secure storage & retention limits
• Hands-on: Build a policy-aware requirements matrix and hazard log (e.g., STPA-lite) for a benign aerial monitoring use case (infrastructure inspection), including geofencing and abort criteria

📅 Day 2 – Robust Perception & Tracking (Safety-Critical & Non-Targeted)

• Robustness concepts: dataset shift, OOD detection, calibration, false-alarm vs miss trade-offs
• Evaluation & validation: test plans, scenario coverage, interpretability & explanations for operator trust
• Privacy-preserving sensing patterns: minimization, masking/redaction, on-device filtering
• Hands-on: Evaluate a pre-trained detector on a public, non-person dataset; compute calibration/OOD metrics, document results in a model card with limitations and safe-use guidelines

📅 Day 3 – Human Oversight, Safe Comms & Multi-Drone Coordination (Benign Missions)

• Human-on-the-loop interfaces: alerts, positive control, rate-limiters, locked-safety modes, no-fly & no-observe zones
• Communications safety: designing fail-safe behaviors for degraded/lost links (no evasion tactics), logging and traceability
• Multi-UAS coordination for public-good scenarios (e.g., search-and-rescue): collision avoidance, deconfliction, fairness of tasking
• Hands-on: Simulate a multi-UAS search-and-rescue mission with strict geofences and lost-link behaviors; produce an assurance case outline (safety case, test evidence, operator SOPs)

• Engineers/ML practitioners, UAS developers, QA/safety, PMs, compliance, public-sector tech.
• Basic Python, core ML (detection/classification), Git/CLI, clear documentation.
• ROS/ArduPilot/DroneSim, STPA/FMEA, model/system cards.
• Commitment to non-weaponized, public-good use; privacy-by-design & human control.