We’re developing proprietary AI vision models and modular, cross-platform autonomy capabilities for compatible unmanned aerial systems, with defense platforms as our primary focus. Our Vision Engine and Autonomy Engine complement one another and can be integrated independently or together. Each capability has its own maturity and evaluation requirements.
Use onboard cameras and IMU measurements to estimate motion and navigate when satellite positioning is unavailable. Depth cameras provide obstacle sensing and local mapping. The current capability supports local, body-relative navigation with an explicit launch reference.
EI2 supports depth cameras, stereo cameras, and LiDAR as sensing options for compatible platform configurations. Sensor selection, interfaces, calibration, onboard compute, and performance are evaluated for the intended aircraft and operating conditions. Support does not imply compatibility with every sensor model or simultaneous sensor fusion in every configuration.
Implemented control and flight-controller integration components are configured around the airframe, mission, and operating conditions. Control approaches are selected and tuned for stability, responsiveness, and compute constraints, then evaluated for the intended flight envelope.
EI2 supplies planning, supervision, and integration components; PX4 supplies underlying low-level stabilization and flight-controller functions. Each integration defines ownership explicitly. Our engine combines EI2 implementations with selected third-party libraries and models; integrated dependencies are not presented as proprietary EI2 inventions.
Developing custom models for object detection, classification, multi-object tracking, identification, and visual scene understanding. Representative training data, held-out accuracy measurements, and edge-inference latency guide evaluation. V0 aerial-object models undergo ongoing training and validation. Our development scope includes drones and Shahed variants, missiles, helicopters, hot air balloons, kites, parachutes, paragliders, and bullets. A category’s inclusion does not establish reliable detection in flight.
Implemented path planning, static and dynamic obstacle avoidance, trajectory planning and control, and navigation and flight-controller integration. Recorded simulation demonstrations show selected configurations in action. Watch demos → AI-based perception is combined with conventional robotics algorithms, planning, and control; not all autonomy components are AI models. Explore autonomous navigation →
Track moving objects, estimate their relative motion, predict potential collisions, and adjust navigation using predicted trajectories and time-to-collision checks. Watch the dynamic avoidance demo →
CPU-only profiles support ARM and Intel GPU-less boards. NVIDIA and Qualcomm compute options support configuration-specific execution; NVIDIA profiles provide acceleration for supported inference, stereo-depth and mapping workloads. Qualcomm support has been tested in SITL; physical hardware testing is in progress. Algorithm and device selection are recorded; unavailable or incompatible configured stages refuse initialization. These options do not imply equal performance, support for every computer, or sustained real-time operation. Evaluate the full sensing-to-command path for the intended platform.
Ten selectable component domains cover sensing, perception, mapping, static and dynamic avoidance, planning, safety supervision, missions, configuration, and diagnostics. Integrate vision alone, static avoidance with its required providers, or a combined configuration. Selection respects dependencies; this is not a public SDK offering.
Choose among implemented planning, stereo-depth, mapping, and trajectory-generation approaches. Selection depends on sensors, compute, aircraft dynamics, and mission requirements. Availability as a component does not establish integrated flight qualification for every combination.
Build local geometry that distinguishes occupied, observed-free, and unknown space. Multiple-view fusion, evidence aging, and GPU mapping support selected configurations. Missing detections do not establish clearance.
Independent supervision checks stale inputs, stopping space, continuous trajectory clearance, command expiry, and control authority. Defined simulation scenarios exercise blocked-route stopping and landing and a specific camera-loss/recovery policy. These checks do not guarantee collision-free flight or cover every failure mode.
Structured telemetry, effective-configuration records, recording and replay, and fault injection support integration and investigation. Source-specific simulation evidence records what was exercised and its limits; diagnostic tools do not establish hardware or production readiness.
Evaluate configured altitude limits, operating boundaries, prohibited volumes, authorization coverage, and terrain-clearance requirements. Current authoritative data and applicable permissions must be supplied and validated. These checks are not regulatory certification or a substitute for operator responsibility.
Implemented components: selectable modules, platform interfaces, compute selection, mapping, planning, supervision, and diagnostics.
Simulation-tested configurations: recorded PX4 scenarios include camera-informed navigation, GPU mapping, blocked-route handling, and a defined camera-loss response. ArduCopter evidence covers limited control only. Not all models, algorithms, and adapters have been tested together.
Still to qualify: physical hardware and flight performance, sustained real-time throughput, and broader operating environments. GPS-denied visual-inertial navigation and dynamic-object collision avoidance are supported software capabilities. Robust degraded-environment operation remains a development objective.
EI2 is a defense technology company developing proprietary AI vision and autonomy for unmanned aircraft. Current engineering includes PX4, ROS 2, simulation-based autonomy development, and computer vision model experiments. Supported autonomy capabilities include GPS-denied camera-and-IMU navigation and dynamic-object collision avoidance. Advanced multi-sensor localization fusion and robust operation in degraded environments remain development objectives. Model and capability maturity vary; production readiness is not established across all use cases.
| Platform | Current scope | Limits |
|---|---|---|
| PX4 / ROS 2 | Primary integration environment; recorded camera-navigation, mapping, and supervision SITL checks. | Configuration-specific simulation evidence; no general physical-flight or real-time qualification. |
| ArduCopter | Limited MAVLink control adapter: supervised velocity commands and observed hold/land behavior, with firmware SITL evidence. | Autonomous navigation and camera avoidance unavailable; other airframes and hardware unqualified. |
| Other vendor platforms | Future adapter development subject to available control and sensor interfaces. | No validated integration or universal compatibility claimed. |
These are development and evaluation objectives, not established performance gains. Each pilot defines a baseline, target metrics, operating conditions, and a compatible hardware configuration. We report measured results with their test scope and limitations.
Measure detection precision and recall, false-alarm rates, tracking continuity, and identity switches on representative, held-out data.
Measure end-to-end perception-to-control latency and processing rate on the target onboard hardware.
Evaluate collision and near-miss rates, minimum obstacle clearance, and avoidance success within defined test scenarios.
Compare mission completion rates, route length, flight time, and human interventions under matched conditions.
Profile compute utilization, memory use, and power consumption for each tested model and hardware configuration.
Track integration effort and validated combinations of flight controllers, sensors, compute hardware, and capabilities.