________________________________________
Hummel Project (Indoor Micro-Drone Prototype) EEG driven
A friend and I are currently building what might be the smallest private mini-drone. Without EMP and camouflage, of course. That was only a thought experiment to check whether those concepts would be feasible at such a small scale. The AI response indicated it would be possible in principle, but not part of this build.
Origin: Back in 2011, during our studies, we ran small flight simulations based on neurofeedback. The current challenge is: can we implement a similar idea using a tiny indoor mini-drone?
Scope / Limitation: Due to its prototype stage, the drone flies indoors only. It is not suitable or designed for outdoor use.
________________________________________
1) Bill of Materials (Components)
Here are all mentioned components, including the latest additions:
1. Frame: Happymodel Mobula6 65mm Frame Kit (14.99 €, 5.5 g)
2. Flight Controller: Happymodel Crazybee F4 Lite 1S AIO (49.99 €, 2.7 g)
3. Motors (4x): RCINPOWER 0702 25000KV (49.90 €, 9.2 g total)
4. Propellers: Gemfan 30mm 3-Blade (8.95 €, 1.6 g)
5. FPV Camera: Caddx Ant Lite (38.00 €, 2.0 g)
6. Video Transmitter (VTX): TBS Unify Pro Nano (42.90 €, 1.5 g)
7. Battery: BetaFPV 300mAh 1S LiPo (9.99 €, 8.5 g)
→ Upgrade: 450mAh (12.0 g)
8. Controller: RadioMaster TX12 Mark II ELRS (119.99 €, 363 g)
→ Optional; lighter alternatives recommended
9. LiDAR Sensor: Benewake TF-Luna (24.90 €, 3.5 g)
10. Optical Flow Sensor: PMW3901 Optical (15.99 €, 2.1 g)
11. Voltage Regulator (BEC): Matek 5V/3A Step-Down (6.90 €, 1.2 g)
12. Capacitor: 470µF Low-ESR (1.50 €, 1.0 g)
13. AI Chip: STM32H7 (~3.0 g) → replacement for ESP32
14. VO₂-Graphene*: ********* (experimental thought, not in build)
15. Piezo-EMP: (DIY ******) (experimental thought, not in build)
Total weight (with 450mAh battery): 43.8 g
Headroom up to 60 g: ~16.2 g
________________________________________
2) Compatibility Check (by Component)
1) Frame (Mobula6 65mm)
• Fits 0702 motors + 30mm props (standard for 65mm whoop frames).
• Space for Crazybee F4 Lite (AIO) and sensors (LiDAR + optical flow).
• Issue: very limited space for STM32H7 (3 g) + Piezo-EMP (1.7 g).
• Solution: flat component layout + lightweight 3D-printed mounting system (~1–2 g).
2) Flight Controller (Crazybee F4 Lite 1S AIO)
• F4 processor (168 MHz) supports PX4 (with porting); good for autonomous control.
• Integrated ESC (1S, 5–12A) matches 25000KV motors + 450mAh battery.
• UART ports for LiDAR, optical flow, and STM32H7.
• Integrated receiver (ELRS) compatible with TX12 or alternatives (e.g., FrSky XM+).
• Issue: limited compute for complex AI (TensorFlow Lite, etc.).
STM32H7 handles AI, but comms (UART/I2C) must be reliable.
• Solution: STM32H7 as AI module over UART; test latency (target max ~10 ms).
3) Motors (0702 25000KV)
• Works well with 1S LiPo (3.7 V) + 30mm props.
• Crazybee ESC supports typical current draw (~6–8A at full load).
• Issue: higher energy demand reduces flight time (approx. 5–7 min with 450mAh).
VO₂-Graphene might help with thermal management, not energy efficiency.
• Solution: optimize with efficient flight mode (PX4 hover tuning / lower RPMs).
4) Propellers (Gemfan 30mm 3-blade)
• Designed for 0702 + 65mm frames; CW/CCW config stabilizes the drone.
• No relevant issues.
5) FPV Camera (Caddx Ant Lite)
• Compatible with VTX (TBS Unify Pro Nano).
• Runs on 3.3–5V; covered by 5V/3A BEC.
• Video can be used for AI processing (e.g., obstacle detection).
• Issue: limited visibility range indoors (2–3m), but sufficient for this use.
• Solution: ROS 2 + OpenCV-based processing.
6) Video Transmitter (TBS Unify Pro Nano)
• 5.8 GHz, 25–400mW; compatible with camera + BEC (5V).
• Issue: power consumption (~0.5–1W) reduces flight time; indoor range is overkill.
• Solution: run at 25mW indoors + add capacitor for stability.
7) Battery (450mAh 1S)
• Works with the 1S Crazybee setup; ~5–7 minutes depending on load.
• Issue: no built-in overvoltage protection; BEC protects 5V rails.
• Solution: keep 470µF capacitor for voltage smoothing.
8) Radio (RadioMaster TX12 Mark II ELRS)
• ELRS matches Crazybee integrated receiver.
• Issue: TX weight irrelevant to drone weight, but setup may be “overkill” for micro tests.
• Solution: either use AIO receiver directly or lighter alternatives depending on the build goal.
9) LiDAR (Benewake TF-Luna)
• UART interface compatible with Crazybee + STM32H7.
• Range 0.2–8m ideal for indoor obstacle avoidance.
• 5V covered by BEC.
• No major issues; integrate into PX4.
10) Optical Flow (PMW3901)
• SPI interface compatible via adapter.
• Good for indoor stabilization on flat/textured surfaces.
• Issue: limited altitude effectiveness (~2m), but fine indoors.
• Solution: use PX4 calibration/tuning.
11) BEC (Matek 5V/3A Step-Down)
• Provides stable 5V for camera, VTX, sensors, STM32H7.
• ~90% efficiency.
• Issue: 3A limit under high overall demand.
• Solution: monitor total draw, reduce VTX power, measure with multimeter.
12) Capacitor (470µF Low-ESR)
• Smooths voltage spikes; protects electronics.
• No issues.
13) STM32H7 AI module
• 480 MHz; supports PX4 and TensorFlow Lite Micro.
• UART/I2C communication possible.
• Issue: limited flash/RAM for complex models.
• Solution: use small models (<20 KB) and test for latency stability (target max ~10 ms).
14 und 15 (experimental thought)
• Light (~g)
• Voltage: (V) ******
• Problem: ********
• Solution: ******
________________________________________
3) Overall Compatibility Summary
• Electrical: components match (5V BEC, 1S LiPo, ESC), but high-load current (~8–10A peak) should be monitored.
• Mechanical: fits in 65mm frame, but STM32H7 + extra modules require tight layout and likely a 3D-printed holder.
• Software: PX4 integrates LiDAR + optical flow + camera; ROS 2 handles higher-level logic. EEG trigger (UART from OpenBCI) must be synchronized.
• Weight: ~43.9g with margin up to 60g.
________________________________________
4) Potential Issues + Fixes
1. Power draw: motors + VTX can drain the battery quickly
→ Fix: low VTX power (25mW), efficient control tuning (PX4)
2. Latency: STM32H7 ↔ Crazybee comms may introduce delay
→ Fix: UART 115200 baud testing + optimize PX4 config
________________________________________
5) Recommendation
Overall, parts are compatible. Key build priorities:
• mount STM32H7 cleanly (3D holder, shielding if needed)
• monitor current consumption and optimize VTX
• test PX4 + ROS 2 + EEG trigger synchronization
________________________________________
6) Improvements Using +10g Budget
1) Larger battery
• Option: 450mAh 1S (~12g vs ~7g)
• Benefit: flight time increases from ~2–3 min to ~5–7 min
• Cost: ~10–12 €
• Weight: +5g
2) Better AI compute module
• Option: STM32H7 w/ AI expansion (2–5g) or a light module like HiSilicon Hi3516 (3–5g)
• Benefit: more headroom for EEG analysis + autonomous logic
• Cost: ~15–20 €
• Weight: +2–5g
3) Additional sensors
• Option: FLIR Lepton module (0.5g) + mount (2–3g) OR second LiDAR (3.5g)
• Benefit: thermal imaging (indoor / story element) or better 360° navigation
• Cost: FLIR ~150 €, LiDAR ~25 €
• Weight: +3–5g
4) More robust structure
• Option: reinforced frame or 3D-printed casing (2–5g)
• Benefit: improved crash stability, easier EEG control (fewer corrections)
• Cost: ~5–10 €
• Weight: +2–5g
Example weight distribution (reinforced build)
• Frame (reinforced): 7.0 g
• FC: 2.7 g
• Motors: 9.2 g
• Props: 1.6 g
• Camera: 1.2 g
• VTX: 0.8 g
• Battery (450mAh): 12.0 g
• Optical flow: 2.1 g
• BEC: 1.2 g
• Capacitor: 1.0 g
• ESP32 (AI): 5.0 g
• FLIR + mount: 3.5 g
• Total: 47.3 g
• Remaining: 12.7 g margin up to 60 g
Costs
• Additional: ~30–50 € (battery + AI module + optional FLIR)
• Total estimate: 364–384 € (excluding FPV goggles)
Benefits
• Flight time: +2–4 minutes
• AI performance: improved EEG processing + autonomy
VO₂ / graphene (not)
________________________________________
________________________________________
(kept as-is)
________________________________________
7) `*****NameOfBrand (Evaluation)
Benefits
1. No-code approach: fast model building without heavy coding
2. Data management: annotation + augmentation (obstacles, surfaces)
3. Cloud processing: train heavier models offboard (useful for indoor tests)
4. Flexibility: supports multiple video sources, edge-device workflows
Relevance
• Autonomy: train obstacle recognition (LiDAR + optical flow + camera)
• EEG control: not directly EEG-native, but EEG could be transformed into visual representations (heatmaps) for training (requires preprocessing)
Challenges
1. Primarily vision-focused, EEG needs conversion work
2. Access/cost unclear after ****** acquisition
3. Edge deployment limits on STM32H7 require heavy compression (e.g., TFLite <20KB)
4. Indoor dataset size may be small (augmentation needed)
Critical view
• Platform may be overkill for a 60g “just for fun” prototype
•
Alternative
• Local training (PyTorch/TFLite) for EEG logic
• Use cloud only for vision (if available)
Conclusion
• Useful for vision training if accessible
• Less suited for EEG directly
• Local workflow may be more practical for this prototype
Next steps
1. Check access to CrowdAI
2. Train EEG logic locally (OpenBCI + small model)
________________________________________
8) Indoor Autonomy Requirements
• Autonomy: obstacle avoidance, position hold, simple tasks (scan room / find target)
• Hardware: STM32H7, PMW3901 optical flow, TF-Luna LiDAR, Caddx Ant Lite camera
• Weight ceiling: 60 g
• Indoor environment: small rooms (e.g., 5×5 m), no GPS
• Programming: complex solutions possible if needed
________________________________________
9) Program + Training Recommendation (PX4 + ROS 2 + RL)
1) Control stack: PX4 + ROS 2
• PX4 for drone control, sensor fusion, autonomy modules
• ROS 2 for middleware + higher-level AI and decision-making
• ROS 2 may run offboard (laptop / optional Pi Zero) while PX4 handles flight control
2) Training approach: Reinforcement Learning (Stable Baselines3)
• RL fits indoor autonomy decision-making (trial-and-error learning)
• Scenario: when no EEG signal, drone switches into autonomous mode
• Train on laptop, deploy compressed model (TFLite) where possible
3) Simulation-first: Gazebo
• Train in simulated indoor environment to reduce real-flight test cycles
• Setup: 5×5 m room with obstacles
• Reward logic: +1 avoid obstacles, -1 collision, +5 stable hover
________________________________________
Detailed Implementation
redacted