Roadmap

Erwin Lejeune — 2026-02-19

Planned improvements, grouped by domain and roughly prioritised within each section. Items marked done have shipped; the rest are open, and most are a reasonable first contribution.

See CONTRIBUTING for the bar a new algorithm has to clear, and CHANGELOG for what has already landed.

Recently shipped

• Fixed-wing aerodynamics rebuilt on the full Beard & McLain coefficient model, with a numerical trim solver, four airframe presets and a derived-gain autopilot.
• VTOL transition with a shared wing model, a rate-limited tilt actuator and a mode-scheduled controller that holds altitude across the handover.
• Reinforcement-learning gym — six environments and a dependency-free trainer.
• flybots CLI with catalogue discovery and a physics self-check.

---

1. Vehicle Models

1.1 Hexacopter / Octocopter Support

• done — Multirotor(MultirotorParams) with configurable arm geometry (X, +, H and coaxial layouts). Quadrotor is a preset over it.
• done — mixer matrices derived from rotor positions and spin directions; the two hard-coded 4×4 matrices are reproduced to machine precision by the derivation.
• done — VehiclePreset.HEX_S550 and VehiclePreset.OCTO_X8.
• Per-rotor failure injection, so a rotor-out on a hexacopter can be flown rather than only reasoned about. The allocation already reports the rank loss; what is missing is a way to kill a motor mid-flight.
• Tricopters, which need a tilting tail servo rather than an extra column in the allocation matrix — an odd rotor ring cannot cancel its own yaw.

1.2 Improved Aerodynamic Models

• Add blade-element-theory (BET) rotor model for more realistic thrust curves at high advance ratios (relevant to VTOL transition).
• Add parasitic + induced drag to the quadrotor fuselage for speed- dependent drag losses visible in high-speed trajectory tracking.
• Add ground-effect thrust augmentation when z < 2 * rotor_radius.

1.3 VTOL Improvements

• done — wing lift as a function of angle of attack and airspeed, shared with the fixed-wing airframe model.
• done — airspeed-scheduled transition, so the rotors only tilt further forward once the wing can take up the slack.
• Back-transition stall protection. The current back-transition deliberately passes through stall while decelerating; the rotors carry the aircraft, but an explicit envelope guard would be better.
• Full tilt-wing model, where the wing itself rotates.
• More than one tilt-rotor preset.

1.4 Fixed-Wing Autonomy

• done — altitude, airspeed and course hold, with gains derived from each airframe's control authority.
• done — mission waypoint navigation: straight-line and orbit vector-field following, half-plane waypoint acceptance, racetrack patterns and return-to-launch.
• L1 adaptive guidance law for path following.
• Wind-aware Dubins path planner.
• Wind and gust models. The aerodynamics already work from an air-relative velocity, so wind enters in one place.

---

2. Control

2.1 Adaptive Control

• Model Reference Adaptive Control (MRAC) that learns unmodelled dynamics (payload shifts, motor degradation) online.
• L1 adaptive controller for disturbance rejection.

2.2 Robust Control

• H-infinity controller synthesis for quadrotor hover.
• Sliding mode controller for aggressive manoeuvres.

2.3 Reinforcement Learning

• done — six environments plus a pure-NumPy trainer (ARS and CEM), and optional Gymnasium registration.
• PPO/SAC agent for aggressive trajectory tracking, replacing the inner loop during acrobatic manoeuvres.
• Domain randomisation over mass, inertia, motor time constants and wind. Policies trained today are tuned to a single airframe.
• Recurrent policies for the partially-observed tasks.

2.4 Unified Controller Architecture

• Merge CascadedPIDController and FlightController into a single configurable stack to eliminate the current dual-stack confusion.
• Provide a ControllerFactory that builds the appropriate stack from a YAML config.

---

3. State Estimation

3.1 Visual-Inertial Odometry (VIO)

• Monocular VIO pipeline: feature tracking → 5-point essential matrix → local BA → EKF fusion with IMU.
• Stereo depth estimation for dense mapping.

3.2 Multi-Sensor Fusion

• Factor-graph-based estimation (iSAM2-style) replacing the current EKF for tighter coupling of GPS, IMU, lidar, and vision.
• Barometer + magnetometer sensor models and fusion.

3.3 SLAM Improvements

• Graph-based SLAM (pose-graph optimisation with loop closure).
• Extend EKF-SLAM to handle dynamic landmarks (moving objects).
• 3D occupancy mapping with OctoMap-style octree compression.

---

4. Perception

4.1 Object Detection & Tracking

• Multi-object tracking (MOT) with Hungarian assignment and Kalman prediction for multiple bounding boxes simultaneously.
• Simulated YOLO-style detector with configurable false positive / false negative rates and detection noise.

4.2 Semantic Mapping

• Label occupancy grid cells with semantic classes (road, building, vegetation, water) for mission-aware planning.
• Height-map construction from lidar point clouds.

4.3 Depth Estimation

• Simulated depth camera sensor model.
• Point cloud registration (ICP) between successive scans.

4.4 Visual Servoing Enhancements

• Image-Based Visual Servoing (IBVS) with proper interaction matrix and Jacobian-based velocity mapping (6-DOF).
• Position-Based Visual Servoing (PBVS) using estimated target pose.
• Hybrid VS (switch IBVS↔PBVS based on feature visibility).

---

5. Path Planning

5.1 Dynamic Re-Planning

• D* Lite for real-time replanning when obstacles are discovered.
• Informed RRT* with path-heuristic focusing.
• Kinodynamic RRT: plan in state space (position + velocity) to produce dynamically feasible trajectories directly.

5.2 Multi-Agent Planning

• Conflict-Based Search (CBS) for decoupled multi-drone deconfliction.
• Velocity obstacles (VO / ORCA) for reactive collision avoidance in dense swarms.

5.3 Energy-Aware Planning

• Battery model (voltage sag under load, capacity vs. temperature).
• Energy-optimal path planning that factors in wind, altitude, and payload.
• Return-to-home with energy reserve constraint.

---

6. Trajectory Planning & Tracking

6.1 Time-Optimal Trajectories

• Time-optimal trajectory through waypoints subject to thrust and tilt constraints (using convex optimisation or iterative methods).
• CPC (Complementary Progress Constraints) for collision-free multi-segment time allocation.

6.2 B-Spline Trajectories

• Uniform B-spline trajectory representation (more numerically stable than polynomial for long paths).
• Online B-spline deformation for obstacle avoidance (gradient-based local replanning à la EGO-Planner).

6.3 Corridor-Constrained Planning

• Safe flight corridor generation from 3D occupancy (convex decomposition).
• Corridor-constrained minimum-snap optimisation.

6.4 MPPI Improvements

• GPU-accelerated rollout sampling (JAX or CUDA kernels) for real-time performance at K>1000 samples.
• Colored-noise MPPI for smoother control sequences.
• Covariance adaptation based on cost landscape curvature.

---

7. Swarm Algorithms

7.1 Task Allocation

• Market-based task allocation (auction algorithm) for heterogeneous missions (inspect, deliver, photograph).
• Dynamic task reallocation when an agent fails.

7.2 Advanced Formation Control

• Bearing-only formation control (no range sensor needed).
• Formation morphing: smooth transition between different shapes.
• Obstacle-aware formation deformation.

7.3 Communication Models

• Realistic communication graph: range-limited, lossy, with latency.
• Consensus under communication delays and packet drops.
• Decentralised SLAM with inter-agent loop closure sharing.

---

8. Environment & Simulation

8.1 Wind and Weather

• Dryden / Von Kármán turbulence model for stochastic wind gusts.
• Steady-state wind field (configurable direction and speed).
• Rain / fog degradation of sensor performance.

8.2 Terrain

• Digital Elevation Model (DEM) terrain for realistic nap-of-the-earth flight.
• Terrain-following altitude controller.

8.3 Dynamic Obstacles

• Moving obstacles (pedestrians, vehicles) with configurable trajectories.
• Pop-up obstacles for reactive avoidance testing.

8.4 Multi-Fidelity Simulation

• Real-time 3D visualisation with PyVista or Open3D (replace matplotlib for interactive exploration).
• Gazebo / MuJoCo bridge for physics-in-the-loop testing.
• Software-in-the-loop (SITL) with PX4/ArduPilot via MAVLink.

---

9. Infrastructure & Tooling

9.1 Configuration System

• YAML/TOML-based simulation configs (vehicle params, controller gains, world definition, viz settings) instead of hard-coded constants in each run.py.
• CLI runner: flybots run --config my_scenario.yaml.

9.2 Benchmarking

• Standardised metrics: RMSE, settling time, overshoot, energy consumption, completion time.
• Automated benchmark suite that runs all sims and produces a comparison table/report.

9.3 Testing

• Property-based testing (Hypothesis) for control stability margins across parameter ranges.
• Fuzzing of sensor inputs (NaN, extreme values, dropouts).
• Regression tests that compare GIF checksums to detect unexpected visual changes.

9.4 Documentation

• Sphinx/MkDocs site with API reference auto-generated from docstrings.
• Tutorial notebooks (Jupyter) for each domain (control, planning, estimation).
• Architecture decision records (ADRs) for major design choices.

9.5 CI/CD

• GitHub Actions: run full test suite + linting on every PR.
• Automated GIF regeneration on main merges (cached with LFS).
• Release automation with changelog generation (commitizen).

---

10. New Simulation Ideas

Domain	Simulation	Description
Control	pid_tuning_comparison	Side-by-side Ziegler-Nichols vs. manual vs. auto-tuned PID
Control	wind_rejection	Hover in turbulence with/without feedforward compensation
Planning	dynamic_replan	D* Lite replanning around discovered obstacles
Planning	multi_drone_deconflict	CBS or ORCA for 4+ drones in shared airspace
Trajectory	bspline_corridor	B-spline through safe flight corridors
Trajectory	time_optimal	Fastest path through gates under thrust constraints
Estimation	vio_demo	Visual-inertial odometry with feature tracks
Perception	multi_target_tracking	MOT with multiple moving ground targets
Perception	semantic_mapping	Labelled occupancy grid from simulated camera
Swarm	task_allocation	Auction-based task assignment for 6+ drones
Swarm	formation_morphing	Smooth hexagon → line → V transitions
Vehicle	payload_delivery	Quadrotor picks up, transports, and drops a payload
Vehicle	autorotation	Emergency landing via autorotation (motor failure)
Environment	urban_delivery	Full mission in city environment with wind + dynamic obstacles

---

11. Code Quality Debt

• [x] Extract magic numbers (world sizes, cruise altitudes, DT values) into shared simulations/common.py module (done: figure_8_ref, line_to_goal, STANDARD_DURATION, frame_indices, COSTMAP_CMAP)
• [x] Add data panels to all simulations (tracking error, speed, stats)
• [x] Standardize double-integrator dynamics across all swarm simulations
• [x] Remove duplicate waypoint_tracking simulation
• [x] Slow down search algorithm visualization with frame duplication
• [x] Fix LQR/geometric/flight_ops broken controllers
• [ ] Unify CascadedPIDController and FlightController stacks
• [ ] Remove ax_side backward-compatibility shim from ThreePanelViz and clean all sims that still reference it
• [ ] Replace fig.add_axes([...]) manual positioning with gridspec everywhere
• [ ] Add type stubs for mpl_toolkits.mplot3d to silence mypy
• [ ] De-duplicate visualization boilerplate across simulation run.py files (many share identical update-function patterns)
• [ ] Add __all__ exports to all __init__.py files
• [ ] Property-based tests for all controllers (stability across parameter ranges)