Fly the approach
Choose Take control, then hold thrust and turn toward the pad. Brake early; land slowly and upright with both legs inside the lit zone. Arrows or A / D turn; W / ↑ fires thrust. Space pauses, R starts a new approach, and M hands over control without restarting. Retry keeps terrain and starting position but uses the latest checked policy.
Try a one-tap engine cut or solar gust, watch the response, then repair or retry. Adjust flight gives finer engine, solar and fuel controls. Engine health changes available thrust, while attitude jets remain independent. A scheduled failure can still happen after a manual edit; repairing it after the failure persists. Solar push overrides the mission’s lateral force until you clear it. Fuel edits drain or refill the tank without counting as fuel burned or learning feedback.
Autopilot sliders adjust seven real controller parameters, not neural-network inputs. Travel speed, descent and clearance affect the route. Thrust and solar compensation match observed forces. Meteor response adjusts avoidance; fuel-saving descent allows extra speed when estimated reserve is low. Revert restores the current checkpoint without rewinding the ship. Save as new agent keeps the original and trains from your tuned policy. Tests of the exact same weights remain valid in this session. Unsaved tuning is cleared when changing agents, setups or reloading.
One cockpit works on desktop and phone. The ship and flight buttons stay visible while you scroll its settings. Each dropdown groups a task; Escape closes it. Mission, ship and terrain changes are staged until Launch setup. Hands-on flights hold their outcome. Clean retry clears interventions and unsaved tuning on the same approach.
Change → train → compare
Every page opening starts a new challenge at round 0 with fresh weights. It never automatically loads a trained pilot. Resume saved is an explicit choice; named agents remain in Agent & saves. Training progress from the current fresh run is autosaved separately. Starting to train another fresh run replaces that last-run autosave, so use Save to file or a named variant to keep an older run permanently.
Your goal is 90% soft landings in each of the six missions. Test all conditions freezes your current tuning (or latest checked weights) and runs 360 fresh approaches: all ships, flat and uneven terrain, five approach groups, and randomized starts. It includes mission engine failures, meteors, solar disturbance and low-fuel approaches. It does not copy live fuel additions or manual engine repairs. Tests do not train or select the policy; changing its weights requires retesting.
A mission passes at 54 of 60 landings. The overall average cannot hide a weak mission. Results show the weakest mission/ship/terrain setup so you can practice it, adjust the policy and try again. Training practices the selected setup; the all-conditions test checks transfer to the others. A pass is an accomplishment within this finite game test, not proof of zero overfitting or success under every conceivable disturbance. No failures or learning delays are scripted.
Choose a landing priority, exploration strength and extra practice group. Fork & apply creates a named copy with the same learned flight policy and your new recipe. Your original agent is unchanged. A changed goal is a new optimization objective, not proof of improvement.
Balanced keeps the original reward weights. Gentle touchdown raises the touchdown-speed penalty. Fuel saver raises the fuel penalty. Fewer commands raises the cost of changing throttle or steering. Landing remains the first criterion when selecting a flight checkpoint; no setting changes fuel capacity, terrain or collision rules.
Narrow exploration makes smaller parameter changes in the tuner; Wide makes larger ones. For PPO it changes action noise at a fork and the entropy incentive during learning. Extra practice biases only training approaches. Every test still covers all five approach groups.
The Train button runs the chosen number of completed rounds, then stops. Stop & compare keeps finished rounds and discards the unfinished optimizer batch. Training pauses while piloting, in a dialog or in a background tab. Compute pace changes effort, not flight playback or physics. Its shortcuts are 1 / 2 / 5 for those speeds, 0 for 10× and 9 for 100×.
Before and after use the same 60 freshly drawn approaches, so a changed test-world sample cannot explain the difference within that run. These tests do not select weights. The second metric shows the selected priority: fuel, touchdown speed or control changes on successful landings. Different policies may land different subsets, so those conditional averages are not perfectly matched comparisons.
Evidence, not a victory animation
Selection check chooses the flight checkpoint, so it is not independent evidence. Unseen tests are separate and never train the policy. Intervals describe sampling uncertainty; 60/60 does not prove universal success or zero overfitting. Repeatedly choosing recipes based on a test also makes it less independent: use Test 60 new approaches for a fresh check.
The plots span all rounds for the current mission, ship, terrain and recipe. Points are actual measurements. Rounds include inherited training when you fork. A dash means unmeasured, not zero. Comparisons last for this session; imported files have their current results recomputed.
What a saved agent contains
A saved agent is a reusable learned policy plus its goal, exploration, practice focus, seed and training state—not just a name. Fork a fuel-saving variant from a pilot you like, keep the original, and compare what each does on the same approach. Save to file downloads JSON; Load opens it locally. Neither sends your file to a server.
Autopilot tuning searches 18 parameters of an engineered controller using CEM. It can improve quickly because the guidance structure already exists. PPO is an experimental neural pilot: 33 physical observations, a 32-unit hidden layer, throttle and steering outputs, clipped updates and a separate value network. It has no teacher or pretrained weights, and some seeds stall.
Half of PPO's eight practice environments use shorter approaches to help discover landings; all evaluations use full mission starts. PPO uses potential-based progress feedback alongside task reward. Reset learning clears weights, optimizer, rounds and history, keeping the mission, recipe and original seed. Keep a backup is selected when there is library space. A reset autopilot still has engineered guidance; it is not a blank neural learner. Older v2 and v3 files are supported.
Simulation notes
Distances, fuel and speed use game units, not calibrated metres. Solar disturbance is exaggerated space-game physics, not atmospheric wind. Collision footprints match the ships. Fuel limits actual engine force. The visible game has no timeout; a test flight still unfinished after its compute budget is reported as unfinished, never relabeled a landing.
Design & learning references