Drive up the ramp, launch upside down and roll upright. Land back at the starting marker, facing the approach.
The wheels drive the ramp, and fan suction keeps the tires in contact. MPC controls six jets from ramp exit to landing.
The dot marks the current position. Each axis scales independently.
Airborne flight predictions
The car is an original glTF model based on reference images. Its dimensions and hardware performance are estimates. The simulation cannot establish what a production Tesla could do.
The layout has four bottom jets and two front side jets. The side nozzles sit at the front body corners, clear of the wheels. Bottom jets push along the car's upward axis. Side jets push inward from either side. There are no roof jets or jets pointing along the car's length. The car must tilt to accelerate or brake along its heading.
The assumed dry mass is 1,428 kg, with up to 250 kg of air. Principal inertias are [800, 2,600, 2,900] kg·m² at full mass. Each nozzle produces up to 8.5 kN at a fixed exhaust speed of 650 m/s. Air use equals total thrust divided by exhaust speed. The model omits tank pressure, cooling, valve delays and recharge.
The wheels drive the approach and ramp. Fan suction keeps the tires in contact. At ramp exit, a browser version of the repository's SCvx and RK4 controller plans the car's motion and orientation by choosing all six nozzle forces. It checks each plan against nonlinear dynamics, applies the first command, then replans from the next measured state.
C-ramp recovery turns the inverted car upright. In the jump scene, choose a landing position and whether to require a 360° barrel roll. Both maneuvers use the same air supply and wheel contacts.
Whole-flight planning optimizes the airborne trajectory from the takeoff state to the selected landing target. Guidance supplies an initial guess and duration; terminal goals, roll gates and clearance constraints shape the optimized path. A nonlinear rollout validates it before the short-horizon MPC tracks it. Amber shows the complete plan and blue shows the current MPC prediction. The duration and contact sequence remain fixed. Planning or MPC rejection stops the run. Jets shut off at physical wheel contact.
The jump scene includes simplified models of a Model 3, Model Y, Semi, Cybertruck and Cybercab. Their dimensions and spacing are approximate. Fixed MuJoCo collision boxes represent the parked vehicles. Clearance is also checked along the planned path. The model does not simulate crash deformation.
MuJoCo simulates the chassis, four rotating wheels, ramp contact, flight and landing at 0.002 s steps. Wheel torques and fan suction drive the ramp. Each jet applies force at its mounting point. Tire friction brakes the car after landing, and body or display-vehicle contact stops the run.
The analytical RK4 model predicts flight for MPC; MuJoCo advances the actual car. The predictor includes wheel spin angular momentum but omits tire contacts. Mass, inertia, nozzle geometry and aerodynamic assumptions remain known. Tires use ellipsoid collision shapes and compliant contacts, with rigid axles and no suspension travel. Tank pressure dynamics, sensor noise and structural failure are omitted.
Run live computes physics and new SCvx plans in a browser worker using Clarabel WASM. A 500 ms sample queue smooths playback while the solver works. The inspector shows how long each solve takes. Compute flight runs as fast as the worker allows and records the result for playback. The dashed blue line shows the prediction; silver shows the path flown. The dot marks the current position, and red marks a rejected plan. The inspector compares requested and delivered thrust. It waits during the approach and ramp, when the wheels control motion. Export saves the states and solver diagnostics.
The car performs an inverted ramp exit and a full barrel roll over the display vehicles. Six jets provide lift and attitude control, fed by a shared 250 kg gas budget.
SCvx MPC DEMOS