AeroHaptix

A uniform ring of actuators around the waist does not give uniform coverage of directions: proprioception is biased. So the layout came from data. Ten participants wore a 46-unit grid (18 on the front, 18 on the back, three on each side of the waist, two on each shoulder, 8 cm apart) and a Meta Quest 2 headset. Each trial buzzed one unit; the participant pointed a controller ray in the direction they felt it from and clicked. 46 units × 5 repetitions × 10 people gave 2,300 position-to-direction pairs.

An MLP (five hidden layers of 64, ReLU) learned the inverse map, from a direction (ϑ, φ) on the sphere to a body position, with a total-variation penalty for smoothness. Then 32 directions were sampled uniformly (ϑ ∈ {30°, 60°, 90°, 120°}, φ every 45°) and pushed through the network. The result puts units on the more curved parts of the body, shoulders and waist, and drops the lowest ring because nobody could feel "below" on the upper body. Toggle the 46-unit study grid in the scene above to see both layouts on one body.

The drone is modelled as a double integrator with state x = [q, q̇] and the acceleration command u as input. Every obstacle b_i gets a barrier function that is positive outside it and zero on its safety boundary:

Because u acts on the acceleration, the barrier has relative degree two and the exponential CBF condition is used:

One linear inequality in u per obstacle. The global safe input is the QP that projects the pilot's u_ref onto all of them at once, which is what earlier CBF haptics rendered as a single force. MultiCBF instead computes a local safe input for each obstacle whose inequality u_ref violates:

The correction's size is the vibration intensity of the actuator whose direction r̂_j best matches the direction to the obstacle. In the released code k₁ = k₂ = 6, K_v = 3 duty levels per m/s² (clipped to the 16 levels), an obstacle is considered within 5 m of its shell, and the two strongest units are kept per frame. The drone tracked the pilot's own velocity command; the cues informed, they never overrode.

Twelve participants (seven with UAV experience) flew a simulated quadrotor through cluttered tunnels in Microsoft AirSim, 70 minutes each. Three feedback conditions in a Latin square: NA, an Xbox controller and no feedback; FSC, a Novint Falcon haptic joystick rendering the single-CBF correction as a force; VSC, the Xbox controller plus AeroHaptix driven by MultiCBF. Three flying directions per condition, each a 5 × 5 × 50 m tunnel with four walls and fifteen cubes, spheres and cylinders placed at random: forward, with the front camera looking where the drone goes; right and upward, where the camera looks +x and the obstacles arrive from the side or from above, mostly unseen.

Measured per flight: total distance, number of collisions, and input disagreement, the mean of ‖u_ref − u_safe‖. Subjectively: NASA-TLX, four Likert questions on sense of control, and a situational-awareness probe: at a random 10 to 40 m the simulation paused and the participant listed the obstacles they perceived, split into visible ones and ones felt only through haptics.

Everything on this page comes from one small Python package and a handful of scripts; there are no trained weights. The package holds the barriers, the exact QP, the released 32-direction table, the tunnel generator, the operator model and the 50 Hz loop; the scripts run the campaign in shards, draw the figures, render the videos and export the scene JSON the 3-D feed plays.

# AeroHaptix/experiments
python -m pytest -q tests                      # gradients vs finite differences, QP vs SLSQP, layout, cues, flights
python run_study.py main   --seeds 24 --shard 0 --nshards 3   # x3 shards
python run_study.py sweeps --seeds 12 --shard 0 --nshards 4   # x4 shards
python run_study.py merge                      # out/study.json + the summary tables
python make_figures.py                         # docs/figures/*.pdf, site PNGs, results JSON, numbers.tex
python make_videos.py --which all              # out/video/*.mp4 -> assets/videos/
python export_web.py                           # assets/data/aerohaptix_scene.json for the 3-D feed
cd ../docs && make                             # aerohaptix.pdf, the technical report

Files. aerohaptix/cbf.py (SuperEllipsoid, Cylinder, Plane, local_safe, global_safe), suit.py (Layout, render_cues, the torso model), world.py (Tunnel), pilot.py (the operator), sim.py (fly). The browser twin is assets/js/aerohaptix_cbf.js; the scenes are aerohaptix_scene3d.js (three.js), the playground aerohaptix_playground.js, the charts aerohaptix_plots.js.

Seeds. Tunnels are seeded per (direction, seed); the operator's noise is seeded per tunnel, so a flight is a deterministic function of (direction, seed, condition) and the same tunnels are flown under every condition.

The paper's own code (AirSim client, Falcon bindings, the ECBF classes and the BLE tactile module) is at huangbj16/DroneSimulation; the actuator direction table on this page is that repository's ActuatorDirections_norm.txt.

Technical report (PDF): the derivations, the solver, the operator model, every figure with its numbers, and the limitations.