Ride the sky with your body

SL-1 Board Design Spec

The board in every image, 3D model and rulebook illustration must be the same board.

A stone smoothed by water, not a skateboard.

Dimensions

80 × 56 × 7 cm (W × D × H). The board travels along its short axis, with the forward long edge as the nose; the rider stands with feet side by side, facing the direction of travel.

Context

Urban Air Mobility (UAM) has moved from science fiction into policy documents. The FAA's Urban Air Mobility Concept of Operations (UAM ConOps) envisages fixed UAM corridors above cities, traffic management supported by third-party service providers, and strategic deconfliction between operators through cooperative rules.

The EU's U-space framework (built on Implementing Regulation (EU) 2021/664) designates U-space airspace for drones and requires services such as network identification, geo-awareness, flight authorisation and traffic information.

Both frameworks are designed almost entirely for "aircraft and operators", not for "an ordinary person standing on a board and steering with their body". How does a rider know which layer they are in? Who yields at a junction? How are rules understood through the body?

FAA UAM ConOps and EU U-space are cited at a conceptual level only; Sky Lane is speculative design and claims no compliance with any specific provision.

Problem Statement

As personal flight becomes an everyday way to commute, existing frameworks remain centred on machines and operators. They lack a body-centred control method and an intuitively legible language of rules.

Sky Lane asks whether the body's centre of gravity as the control interface, and layered lanes plus a visual grammar as the rule interface, can let ordinary people share the city sky safely and calmly.

Prototype input: camera pose estimation

MediaPipe Pose Landmarker (@mediapipe/tasks-vision) outputs 33 body landmarks per frame, each with x and y (0–1 coordinates relative to the frame), z (relative depth, unstable) and visibility (0–1).

Scale normalisation: every distance is divided by the calibrated shoulder width (SW) or torso length (TL = distance from shoulder midpoint to hip midpoint), so thresholds stay the same whether the rider is near or far, tall or short.

Centre of gravity (CoG): a weighted average of the shoulder midpoint (0.4) and hip midpoint (0.6). Not biomechanically exact, but stable enough for side lean and crouch.

Latency budget: total camera-to-screen latency target < 150 ms (pose estimation approx. 30–60 ms + filtering approx. 30 ms + rendering approx. 16 ms).

• Response curve From the dead zone to full scale, an ease-in curve: output = input to the power of 1.6. Novice mode uses 2.0 (steadier), expert mode 1.3.
• Hysteresis Discrete actions use an entry threshold above the exit threshold: crouch engages at 0.12 × TL and releases only at 0.08 × TL, preventing jitter at the boundary.
• Smoothing One Euro Filter (Casiez et al., 2012): stronger filtering during slow movement, lower latency during fast movement. Initial parameters minCutoff = 1.0 · beta = 0.02 · dCutoff = 1.0, tuned by self-testing. Fallback: EMA (α = 0.25).

Other accessibility

• Keyboard / gamepad fallback WASD or arrow keys + Space (hover) + Q/E (climb / descend); also for quick trials and debugging
• One-sided mode Hover and climb / descend with one arm only
• Colour-blind safe Every state is encoded by shape and rhythm as well
• Reduced movement Turn on cruise assist and move only to make changes
• Privacy Camera frames are processed locally in the browser and never uploaded; the demo's start page states this clearly

Junction protocol · vertical separation + turning roundabout

• 1 Through split 150 m before the junction, the layer splits into upper and lower halves: north–south through traffic takes the upper half (e.g. 39–45 m on L2), east–west the lower half (30–36 m). They never cross, so through traffic never waits.
• 2 Turning roundabout Turning boards enter a right-hand turn lane 100 m before the junction and merge into the ring lane at mid-layer (50 m diameter, one-way anticlockwise).
• 3 Entry rule Traffic in the ring has priority; steady blue light ring = merge, slow orange pulse = slow and watch, fast orange flash = stop and hover.
• 4 Exit Return to the matching through half-layer; full-layer lanes resume after 150 m.
• 5 L3 Very few junctions, fully grade-separated; to turn, descend to L2 first.
• 6 L1 Small roundabouts (24 m diameter), no vertical split.

State machine: APPROACH → SPLIT → YIELD_CHECK → IN_RING → EXIT → CRUISE

Take-off: step on → calibration check (recalibrate if > 24 hours since the last) → request a slot → pad light steady blue → rise / raise a hand to climb vertically to the cone top (1 m/s) → find a gap and merge into L1 (priority 6)

Emergency Priority

• 1 The emergency board creates an emergency corridor 600 m ahead along its route (8 m wide, on the side next to the central median).
• 2 Riders in the corridor perform the standard yield manoeuvre: shift right to the outer lane edge → descend 3 m → hover.
• 3 For anyone who has not complied within 3 seconds, the system takes over and completes it (the HUD shows "Yielded automatically").
• 4 Emergency boards can drop from layer E into any layer; riders within 200 m either side of the entry point yield in advance.
• 5 The corridor clears 5 seconds after the board passes; riders resume in order, staggered 1 s apart, to prevent rear-end collisions.
• 6 Emergency board at a junction: every roundabout entry flashes fast orange; riders in the ring leave at the nearest exit and hover.

Safety Score: starts at 100 · F2 −5 · F3 −10 · F4 −25 · collision −30; also one of the self-test metrics

Every 0.1 s · each AI rider and the player's assist system

• Perceive Every board within 80 m around and 10 m above or below.
• Predict Assuming constant speed and heading, compute the time of closest approach within the next 4 seconds (TCPA) and the distance at that moment (DCPA).
• Judge A conflict exists when that distance is below the safe radius (3 m, novices 4.5 m) and TCPA is within 4 seconds.
• Decide By the right-of-way list, the lower-priority party yields; at equal rank, the rider from the right and the rider being overtaken have priority.
• Resolve In order of cost: slow down → shift laterally within the lane → move to another lane in the same layer → hover → change layer (only inside a climb spiral or in an emergency). Choose the first manoeuvre that removes the conflict.
• Recover 2 seconds after the conflict clears, return smoothly to the original lane and target speed.

Rule-based agents + conflict prediction: easy to explain and debug; velocity obstacles / ORCA could serve as an optional comparison.

Pseudo-code

every 100 ms, for each rider R:
    neighbors = riders within 80 m horizontally and 10 m vertically of R
    conflicts = []
    for each N in neighbors:
        relPos = N.position - R.position
        relVel = N.velocity - R.velocity
        if |relVel| < 0.01:
            tcpa = 0
        else:
            tcpa = - dot(relPos, relVel) / dot(relVel, relVel)
        if tcpa > LOOKAHEAD: continue             # LOOKAHEAD = 4 s
        tcpa = max(tcpa, 0)                       # already diverging -> now
        dcpa = | relPos + relVel * tcpa |
        safeR = SAFE_RADIUS(R) + SAFE_RADIUS(N)   # 1.5 m each, beginner 2.25 m
        if dcpa < safeR:
            conflicts.append({N, tcpa, dcpa})

    sort conflicts by tcpa (most urgent first)
    for each C in conflicts:
        if mustYield(R, C.N):                     # right-of-way table
            action = firstWorking([
                slowDown(R, toAvoid=C),
                shiftWithinLane(R, side=RIGHT, max=1.5 m),
                changeLaneSameLayer(R),
                hover(R),
                changeLayer(R) if R.inSpiral or EMERGENCY
            ], test = predictedDcpa(action) >= safeR)
            R.apply(action); log("YIELD", R, C.N, action)
            break
    if no conflicts for 2 s:
        R.recoverTo(laneCenter, targetSpeed)

function mustYield(R, N):
    if priority(R) != priority(N): return priority(R) > priority(N)  # larger number = lower priority
    if headOn(R, N):    return true                # both shift right
    if overtaking(R, N): return true               # overtaker yields
    if N is on R's right: return true              # right-hand rule
    return false

Logging

Each avoidance writes one log entry, exported from the browser as CSV, as the data source for stress tests:

time, riderId, otherId, layer, tcpa, dcpa, action, resolved(bool)

Collisions (board centres < 1.0 m apart) and near misses (< 2.0 m) are logged separately.

Principles and accessibility

Supporting neutrals (rulebook and HUD only): mid grey #8A8F9C (secondary text, inactive states) and pale blue-grey #DCE6EE (layer fills at 40% opacity).

If orange exceeds 5% of a frame, the frame is too tense or orange is being overused.

Sky blue has low contrast against off-white, so blue never carries text; all text is deep navy. Every blue/orange distinction is also marked by shape or rhythm.

At most one orange element holds the visual centre at a time · HUD area < 12% of the screen when cruising · all figures in monospace