Controls, planners, and diagnostics

The laboratory has one executed world and a collection of possible futures. Step asks the controller to examine those futures and apply its decision to the executed world: one action for FMC, or a selected trajectory for Wave Jump (one action if all final walkers are dead). The bright vehicle shows what happened; the paths and cloud show what the planner considered. Keeping those two roles separate makes the controls much easier to understand.

Start with Getting started with the control laboratory for installation and a first run. This page is the operating reference. Scenes, agents, and the editor explains the tasks and editor, Record, replay, and continue experiments explains recordings, and Experiments, comparisons, and performance explains controlled comparisons.

Mobile layout

On phones, the environment occupies most of the screen, with Run, Step, and Restart always accessible. Secondary panels start closed and overlay the world. Opening a panel does not pause execution or change the recording.

Entry point	Controls
Menu	Tasks, Settings, Save / Open
Tools	Inspect, Edit, Drive, camera controls, Experiments, Inspector
Replay	World motion, planner decisions, events and notes

Use a panel’s Close button or Escape to dismiss it. Only one panel opens at a time. Short landscape windows on touch devices use the same compact layout; desktop keeps the expanded controls.

Choose a task, run, and inspect a Wave

Wait until the backend indicator reports WEBASSEMBLY and the run buttons become available. A fresh scene starts paused at tick zero; no action runs automatically. The first-visit chooser introduces the six tasks; Tasks opens it again later. Choose a task, press Run, then click a vehicle to inspect the result. The short introduction can be dismissed, and the browser remembers that choice.

The toolbar keeps the environment, execution controls, and status above the world. Setup, Controller, Rewards, and View organize the settings. Advanced sections start collapsed and remember their expanded state. On narrow screens, Settings and Inspector open drawers so the world remains the main surface.

Changing a physical or planner setting edits a draft. The pending-changes summary shows the old and proposed values; Apply and restart commits them together in one rebuild. The active world continues to use its existing configuration until then. Discard changes restores that configuration. Nonempty runs are saved on this device before replacement; a failed save keeps the original run available and offers recovery choices. See Record, replay, and continue experiments for storage and exports.

Flight is a three-way scene choice hiding behind one field. If environment.flight is omitted (or null), the lab detects flight when at least one controlled body is marked flight_capable; setting it to true forces flight on, and setting it to false forces it off, even for a flight-capable body. In flight, the environment applies a downward acceleration of 9.81 m/s² by default (the environment.downward_gravity value). A rocket or drone must therefore spend some of its thrust countering gravity before that thrust can produce upward acceleration.

Control	Operation	What to expect
Environment	Choose a preset from the scene catalog.	Stages the chosen scene; Apply and restart loads it paused after preserving the current run.
Flight override (environment.flight)	Omitted/null, true, or false	Omitted/null auto-detects flight from a controlled flight_capable body; true enables flight; false disables it.
Downward gravity (environment.downward_gravity)	9.81 m/s²; scene value	Downward acceleration used in flight mode. A rocket or drone’s propulsion must counter it to climb or hover.
Die on wall collision (physics.lethal_walls)	Off in all shipped presets; under Setup → World physics	When enabled, any controlled vehicle touching an outer wall or hole boundary ends the whole world. Apply and restart commits this physics setting.
Vehicle type	Choose Rockets, Drones, Karts, or Harvesters in any environment or racing track.	Stages physics and visuals for all controlled vehicles while preserving other world edits. Apply and restart commits the fleet change paused. Each preset keeps its default until you choose a type.
Vehicle count	Set the vehicle count to a whole number from 1 to 128 in any environment.	Starts from the preset’s count. Stages a fleet change for Apply and restart; invalid entries leave the active scene intact.
Problem properties	Set each controlled agent type’s action multiplier from 0× to 10×, independently for every compiled degree of freedom.	1× keeps the native range; 0× disables a channel; larger values expand the action range and its built-in physical output. Apply and restart commits the draft values to the scene’s agent type properties and rebuilds paused.
Run	Repeatedly plan and execute actions using the selected clock.	Button becomes Pause.
Pause	Stop further execution.	The displayed state remains available for inspection and export. Wave Jump preserves its remaining trajectory for resumption.
Step action / Execute trajectory	Plan once, then execute one action, or the selected trajectory for Wave Jump; if all final walkers are dead, execute only its first positive-duration action.	Pauses continuous running and waits for planning with either clock. A paused Wave Jump trajectory finishes its remaining actions.
Restart	Rebuild the active scene with its active seed.	Preserves the previous run, resets task progress, and starts paused. With pending edits the button reads Apply and restart and commits those edits together.
Start driving / Step physics frame	In Drive mode, start continuous physics or apply the current command for one frame.	Start driving becomes Pause; releasing keys returns to neutral input while motion continues.
Advance Wave population	Advance the native FMC population by one Wave iteration.	Displays population row zero as the world, with a labeled Wave selection recording cut.

The vehicle controls appear in Setup in all six environments. The tab remembers your chosen type separately for each environment, with one shared choice across racing tracks. Changing type keeps the vehicle count and starting positions, world edits, rocks, tethers, rewards, and environment settings. An explicit Flight mode choice remains in effect; automatic mode detects flight for rockets and drones. Restart retains the current scene.

Imported scenes preserve explicit settings; omitted fields receive their defaults, including a wall collision penalty of 100. An explicit physics.lethal_walls: true in an older import remains enabled. A mixed or unrecognized fleet displays the disabled Mixed / custom placeholder; choosing a standard type replaces the whole controlled fleet. Vehicle count counts physical vehicles; Walkers counts planner candidates, each representing a possible future for the whole group.

Mining environments expose Rock size from 0.1× to 2× and Rock weight from 0.01× to 10× in their rock settings. Weight can make a rock one hundred times lighter or ten times heavier without changing its hull or rendered size. They also expose Hook stiffness (N/m). The numeric input accepts values from 0 to 1,000,000 and stays synchronized with a logarithmic slider. These settings update the draft; Update draft also stages the displayed rock properties. Press Apply and restart to commit the value to every tow hook and leave the new world paused. Pending edits do not change the running physics. The presets retain their stiffness defaults: 35 N/m for collaborative mining and 25 N/m for harvesting.

Beside the rock properties, the Flight mode control makes that scene choice visible without editing JSON. AUTO preserves capability-based detection: flight is enabled when a controlled body is marked flight_capable. From AUTO, the first click forces the opposite of the current effective mode: OFF for an auto-detected rocket/drone scene, or ON for an auto-planar scene. Subsequent clicks toggle the forced ON/OFF state. Either forced choice enters the draft. Apply and restart rebuilds the scene and leaves the new world paused. The control changes the mode for the rebuilt scene, not the already-running physics.

A low stiffness makes a hook stretch like a rubber band. Raising it makes the connection approximately fixed in length, but it remains a spring with finite stiffness. It can still swing: resisting changes in length does not stop a rocket moving around the rock. At 0, the spring force disappears while radial damping remains, resisting relative motion along the hook. Zero therefore does not detach the hook or disable all of its forces.

Advance Wave population is an inspection tool for the native Fractal Gas search. Its first click initializes a population from the current world; later clicks advance that same population. It runs native FMC even when another controller is selected. It does not execute the usual mean first action in the original world. The displayed row is one population member, not a declaration of the best future. Wave runs in the serial simulation worker; Worker threads controls the separate live planner. Selected-action risk is not evaluated in this mode. This inspection button is separate from the Wave Jump controller, which executes the chosen branch through ordinary physics steps in the original world.

Use Restart before switching from a Wave demonstration to an ordinary control trial. Wave checkpoints can preserve its population; see Record, replay, and continue experiments. The authoritative world stops continuous running when its terminal flag is set. A collision only ends an episode if that scene’s physics and task rules make it terminal. All shipped presets start with Die on wall collision off and a wall collision penalty of 100. When wall death is enabled, contact by any controlled vehicle ends the whole world, and the wall penalty is still charged on that death frame. Passive cargo and hooks cause neither wall reward penalties nor wall-contact death. Retained-rock physics is unchanged; mining retains its 64-action horizon.

Simulation time and parallel execution

A physics frame has duration scene.physics.dt seconds, with 1/60 second used by the browser when it is omitted. Action frames is an integer count of these frames. With the defaults, one action lasts 6 / 60 = 0.1 simulated seconds and a 16-action horizon reaches nominally 1.6 seconds ahead. Terminal worlds can stop earlier. Increasing the horizon spends more computation looking ahead; it does not change the physics step size.

Reproducible · wait for planning holds the executed world still until the requested search finishes. It then applies the action for the configured duration. Simulation can run faster or slower than wall time. Use this mode to compare decisions without making CPU scheduling part of the control problem.

Real time · fixed simulation clock, for controllers other than Wave Jump, advances the world on a scheduled physics clock while the planner works separately. The planner starts from a prediction of the next action boundary under the current action. A result is accepted only when its scene revision, target tick, and complete root snapshot match the executed world at that boundary. If no acceptable result arrives, the simulation uses the neutral action and increments MISSED DEADLINES. Neutral means zero clamped to each channel’s valid bounds; it need not cancel velocity, gravity, or momentum.

The search budget is the action duration in milliseconds minus 20 ms, with a minimum of 1 ms. The planner checks this budget between incremental advances; one advance and the subsequent risk probes can overrun it. The simulation catches up at most four physics steps per timer callback, then reschedules if still behind. Browser throttling therefore prevents this mode from being a hard real-time guarantee.

Wave Jump waits for the complete search with either clock: the executed world stays still while it plans. Real-time mode paces only the selected trajectory’s execution. Each action uses its recorded edge duration, which can be shorter than Action frames when a sampled world terminates. Zero-duration edges are skipped. The trajectory duration is the sum of these frame counts times scene.physics.dt.

Session control	Default and range	Meaning
Clock (clock)	reproducible; alternative realtime	Worker scheduling policy. A change is staged until Apply and restart.
Worker threads (threads)	8; integers 1–64	Total native simulation threads in the live planning engine, including its calling thread.

The browser prewarms the selected pthread pool: 64 total threads means up to 63 additional pthread workers. The authoritative world, prediction world, risk batch, and Experiments worker remain serial. JavaScript sampling and optimizer updates also run in their planning worker rather than across those native threads. More threads help only when there is enough batch work and hardware capacity. Small populations can spend more time coordinating than simulating.

Threaded planning requires the pthread build and cross-origin isolation. If either is unavailable, the planner falls back to one thread. Read the masthead’s actual thread count after loading; the requested number alone does not establish that parallelism is active. Build instructions are in Getting started with the control laboratory. Clock and thread count are session execution controls, not members of the current portable planner-settings object. Set them explicitly when reproducing a run.

Shared settings and FMC controls

The Controller panel holds the controller choice and common planner settings. Controller, seed, clock, thread count, and Tree changes enter a shared draft. Editing Walkers and Horizon, for example, produces two pending changes and one rebuild when you press Apply and restart. Recordings and the running planner continue to use the active values while you edit. Invalid values must be corrected before Apply can replace the world; Discard changes restores the active values.

Reward-only changes use Apply to current run, described below. If the draft also contains changes that require a restart, Apply and restart commits the whole draft together. Presentation controls in View take effect immediately. Algorithm-specific values are remembered while switching controllers within the current page session. The table gives browser defaults and UI ranges; lower-level APIs can have different limits.

UI label	Settings key	Default; UI range	Meaning and applicability
Controller	algorithm	wave-jump	fmc, wave-jump, random, cem, icem, or mppi.
Walkers	walkers	128; integers 1–8192	FMC/Wave Jump population or shooting batch capacity. Random control does not use a rollout population.
Horizon	horizon	64; integers 1–4096	FMC/Wave Jump normal search depth; Wave Jump can extend it when shared-path execution is enabled. Action depth per shooting round. Ignored by random action selection.
Stop at first bifurcation	consensus_prefix	Checked (true)	Wave Jump only: execute the recorded ancestral path shared by every alive final walker, stopping before their branches diverge.
Maximum search horizon	max_horizon	0 (automatic); integers 0–4096	Wave Jump shared-path mode only. Zero means twice Horizon, capped at 4096. An explicit nonzero value must be at least Horizon.
Action frames	frames	12; integers 1–60	Physics frames per candidate action. Wave Jump executes each selected edge for its actual recorded duration; other controllers execute one action for this count.
Seed	seed	7; integers 0–4294967295	World reset seed and base planner seed; successive decisions derive seeds by adding the decision count modulo 2^32.
Diversity coefficient (in Reward terms)	distance_coef	1; 0–10, increment 0.1	FMC exponent on rescaled observation distance in cloning fitness.
Reward coefficient (in Reward terms)	reward_coef	1; 0–10, increment 0.1	FMC exponent on rescaled reward signal in cloning fitness. Does not edit scene reward weights.
Action noise	noise	0.2; 0–10, increment 0.05	FMC Gaussian standard deviation in each channel’s action units when perturbing inherited actions.
Elites	elites	0; integer 0–Walkers	FMC elite-bank size. The UI initially caps at 128 and updates that cap when Walkers changes.
Perturb inherited actions	inertial	Checked (true)	FMC: after the first iteration, perturb the selected companion’s inherited action; unchecked samples fresh uniform actions every iteration.

Fractal Monte Carlo repeatedly compares population members using observation distance and accumulated reward, copies selected companion worlds, and advances their actions. The first actions travel with their descendants. At the end, the controller averages those inherited first actions over the final population; cloning supplies the implicit weighting. If no population members survive, it returns the neutral action. It does not simply choose the highest-reward leaf.

Wave Jump uses the same population search parameters and cloning procedure as FMC. Stop at first bifurcation is checked by default, including when saved settings omit this option. An explicitly saved unchecked setting remains unchecked. With Stop at first bifurcation unchecked, after the search, it selects the alive final walker with the highest accumulated path reward; ties choose the lower walker index. Alive means nonterminal according to the native physics. It follows that walker’s recorded parent links back to the root and executes the resulting action sequence in forward order. This follows the ancestry through cloning, rather than reading successive actions from one walker slot. Selection always uses accumulated reward, independently of any reward-signal setting used by resampling.

Imagine the search finds a useful sequence of turns. FMC uses its population to choose the next turn, then searches again. With the toggle unchecked, Wave Jump commits to the whole selected sequence before it searches again. If every final walker is dead, it selects the highest accumulated-reward final walker, with the same tie rule, but executes only the first positive-duration action for its recorded frame count before replanning. It therefore makes one planning decision per executed trajectory, including this one-action fallback. The best final walker need not be the best node ever sampled. Execution stops if the actual world terminates. An empty executable path stops with a status message instead of starting repeated searches.

With Stop at first bifurcation checked, trace the ancestry of every alive final walker back toward the root. Execute their shared initial chain and stop before the first branch where these surviving futures disagree. Agreement means the same recorded ancestors; two independently sampled edges with similar actions do not count as agreement. Discarded branches and archived elites outside the current population do not enter this comparison. A single survivor shares its entire path with itself, so its full path is executable.

The search first reaches the normal Horizon. If the shared chain contains no positive-duration action, it continues the same population search one iteration at a time, stopping as soon as an executable shared prefix appears. Maximum search horizon bounds this extension: zero chooses twice the normal horizon, capped at 4096; a nonzero value must lie between the normal horizon and 4096. If agreement is still absent at that limit, execute just the best surviving path’s first positive-duration action for its recorded duration. If all walkers die, use the highest-score single-action fallback immediately. Either fallback searches again after that action unless the actual world terminates. The world stays still throughout the search and any extension. Step completes one search and its shared prefix or fallback, then pauses.

Pause retains the action index and remaining frames. A planner checkpoint preserves these alongside the world and search state, so restoration can resume either a search, including its extension, or a partly executed trajectory without repeating completed work. Checkpoints retain the shared-path setting and effective search limits; legacy in-flight planner checkpoints without a saved shared-path setting restore with the toggle off to preserve their original execution. Changing the scene, applied rewards, algorithm, or world state discards the queued trajectory.

The two coefficient fields are exponents in a product of rescaled distance and reward signals, not an additive meter of distance plus points. Setting an exponent to zero removes that factor’s variation from this product. Elites retains high-reward states together with their actions and ancestry for reinsertion into later Wave iterations. It is separate from a shooting controller’s elite fraction.

The coefficients remain visible in Reward terms and the other FMC-specific controls remain under Planner settings when another algorithm is selected, but random, CEM, iCEM, and MPPI do not consume them. In particular, changing Action noise does not change iCEM or MPPI exploration. Their own noise fields are listed next. The algorithm-engine contract is described in Engine architecture and extension guide and Continuous-control laboratory.

Reward coefficients and term weights

Reward terms is expanded by default. Diversity coefficient and Reward coefficient each have a synchronized slider and numeric input. They control how FMC and Wave Jump select possible futures. The term weights beneath them control the reward earned by the simulated world: movement, target progress, collisions, pickups, deliveries, checkpoints, formation, and full loads. Those reward weights also affect the futures evaluated by the shooting controllers.

For reward-only edits, press Apply to current run. This applies the coefficients and term weights together. At a physics-frame boundary, the lab prepares replacement native engines and a planner, discards old plans, and carries across the current world state, tick, cargo, camera, selection, and decision count. A running experiment resumes; a paused experiment remains paused. If preparation fails, the existing world and settings are retained.

Think of the recording as a continuous film with markers, not a stack of unrelated films. Applying a reward configuration adds a configuration and snapshot boundary; it does not rewind the world or recalculate earlier rewards. Replay continuation restores the historical reward configuration at each such boundary, so the old frames keep the rewards they actually recorded. Typing a pending value therefore changes neither the running physics nor the existing recording until Apply to current run succeeds.

Distance travelled² defaults to weight 1; set it explicitly to 0 to disable the movement bonus. At each physics frame, each controlled vehicle contributes its squared displacement, Δx² + Δy², in square metres. The reward is the weight times the mean of these contributions across the vehicles. Cargo bodies do not contribute, stationary vehicles contribute zero, and respawn teleportation does not count as travel. These frame rewards are summed over an action or journey; the total journey distance is not squared. Movement in any direction earns this bonus, while Target progress separately rewards approaching the task target.

For Tandem flight, Formation reward (rewards.formation) defaults to 50 and accepts values from 0–100. Each physics frame, it pays this weight times the product of the distance-agreement scores for all controlled-body pairs. The engine measures centre distances after motion and before scene mechanics and respawns. A stationary group can earn this reward, and setting Checkpoint proximity to zero leaves it active. Set Formation reward to zero to disable it. Use Edit complete scene JSON to choose the distances; see Scene JSON reference for the formula and formation_pairs.

Tandem flight labels rewards.progress as Checkpoint proximity and defaults it to 1. It also defaults the checkpoint bonus rewards.gate = 30, Distance travelled² = 1, Wall collision penalty = 100, and body collision penalty 2. All other reward terms, including Hooked rock travel and cargo.full_reward, default to 0. Explicit custom weights remain effective. Reset defaults stages these values for Tandem flight, including Formation reward = 50; press Apply to current run to apply them.

Tandem checkpoints advance together in the authored gate order. At the start of each physics frame, the smallest controlled-body checkpoint counter sets the shared stage. Only vehicles at that stage can register a crossing. A vehicle that has already crossed waits and receives no further crossing credit while the others approach. When everyone has crossed, the next checkpoint unlocks on the following frame. Formation and movement rewards remain active while a vehicle waits.

Checkpoint proximity pays on every physics frame, including while vehicles that have crossed are waiting. Let r be the shared active checkpoint’s radius and let d̄ be the mean distance from that checkpoint to all controlled vehicles. The term pays rewards.progress × r / (r + d̄). Thus staying near the checkpoint continues to earn a positive reward. Moving farther away reduces the reward toward zero, but never makes it negative. Cleared and waiting vehicles remain in the distance mean.

Each eligible crossing pays rewards.gate divided by the total controlled count: with two vehicles and bonus 30, each receives 15. Setting either reward weight to zero disables its contribution while preserving the existing counters and shared checkpoint lock. See Scene JSON reference for the exact reward definition.

Rewards → Wall collision penalty (rewards.wall_collision) accepts 0–10,000 and defaults to 100 in every Control Lab task, including harvesting, mining, and older imported scenes that omit the field. Set it to 0 to disable the penalty. Each controlled vehicle touching an outer wall or hole boundary is charged once per physics frame. Sustained contact costs the penalty again on every frame; corners, multiple contact points, and physics substeps add no extra charges within that frame. Use Apply to current run to change this reward while preserving the current world state. Wall-contact death is controlled separately by Setup → World physics → Die on wall collision, which requires Apply and restart. Only applied wall penalty and wall-death options are remembered per named environment when switching presets in the current session; abandoned drafts are not remembered, and fresh tasks restore preset defaults.

The existing rewards.collision weight now applies only to vehicle/body contacts. Harvesting still disables body collisions; its allowed reward terms are progress, distance_squared, catch, and wall_collision.

Hooked rock travel defaults to 0 for Tandem flight and 1 reward per metre for other tasks. Its slider runs from 0–10 in increments of 0.1; the numeric input accepts 0–1,000. Set it to 0 to disable this term, then press Apply to current run to apply the change while preserving the physical world. At the start of each physics frame, the engine identifies cargo rocks hooked to an active controlled vehicle. It sums their distances travelled during that frame, sqrt(Δx² + Δy²), and multiplies the sum by this weight. Each distinct rock counts once, even when several hooks hold it. Moving a rocket around a stationary rock earns none of this bonus, and respawn teleportation does not count as travel. Rock motion in any direction earns it, including circular motion; Target progress and Delivery bonus provide the incentive to bring the rock to the refinery.

In mining, Delivery bonus pays for bringing a rock into the refinery; Target progress rewards approaching a rock and, once attached, moving the hauled rock toward the refinery. Progress now measures both ends of each physics frame against the hauling target selected at that frame’s start. Breaking a hook therefore cannot earn a bonus simply by switching the distance being measured from refinery distance to rocket-to-rock distance. This removes an incentive for repeated attachment and breakage; it does not prevent physical swinging around a rock, or the movement bonus when its weight is positive.

Term weights are saved with the scene; the FMC coefficients retain the planner settings keys distance_coef and reward_coef. A scene that omits rewards.distance_squared receives the default weight of 1, including older scene files. An explicit zero remains disabled. See Scene JSON reference for the scene fields.

The hooked-rock weight is saved as rewards.hooked_rock_distance. Omitting it uses the task’s default above; an explicit 0 remains disabled.

For older scenes that omit rewards.wall_collision, the new default affects future and resimulated rewards. Historical stored records are not rewritten, and the snapshot layout is unchanged.

Random, CEM, iCEM, and MPPI

Seeded random baseline samples one independent uniform value within each channel’s declared bounds. It performs no search before executing that joint action. Its live risk diagnostic still simulates continuations, so selecting random does not eliminate all diagnostic computation.

Shooting controllers sample whole action sequences, simulate them from the same root, and use their summed rewards to revise a distribution. They execute only the first action and plan again from the resulting world. Search rounds counts distribution updates, while Horizon counts actions inside each sequence. Equal walkers and horizon therefore need not mean equal work across algorithms.

UI label	Settings key	Default; range	Applies to
Search rounds	search_iterations	3; integers 1–128	CEM, iCEM, MPPI.
Elite fraction	icem_elite_fraction	0.1; 0.01–0.5	iCEM: max(1, floor(Walkers × fraction)) elites.
Reuse elite fraction	icem_keep_fraction	0.3; 0–1	iCEM: fraction of retained elites offered to the next batch.
Population decay factor	icem_decay	1.25; 1–10	iCEM: reduce active candidates across rounds, within population and elite constraints.
Noise spectral exponent	icem_beta	2; 0–4	iCEM: exponent of the 1/f^beta noise spectrum; larger values emphasize slower temporal variation.
Distribution momentum	icem_alpha	0.1; 0–0.99	iCEM: fraction of the previous distribution retained in mean/std updates.
Initial normalized noise	icem_sigma	0.5; 0.001–2	iCEM: initial standard deviation in normalized action coordinates.
Minimum normalized noise	icem_min_sigma	0.01; 0.0001–1	iCEM: standard deviation floor; must not exceed icem_sigma.
Temperature (reward units)	mppi_temperature	1; 0.000001–1000000	MPPI: scale for reward differences in exponential weights.
Normalized exploration noise	mppi_sigma	0.5; 0.001–2	MPPI: fixed diagonal Gaussian standard deviation.

Cross-entropy shooting starts each decision with a channel-midpoint mean and half-channel-range standard deviation. It samples independent clipped Gaussian sequences, retains the top 15% (at least one), and refits their mean and standard deviation, with a floor of 0.03 in channel units. It returns the first action of the best sequence in the latest completed round. This basic CEM resets its distribution at every decision.

iCEM · improved cross-entropy adds temporally correlated noise, a shifted previous mean and elites, elite reuse, distribution momentum, and decreasing active population across rounds. Retained sequences are evaluated again from the current root. The final round also evaluates the mean. Reused elites and that mean occupy slots inside Walkers, rather than adding extra worlds. The returned first action comes from the best evaluated sequence across completed rounds of this decision.

MPPI · path integral control shifts its previous optimized sequence, samples a fixed Gaussian around it, and updates the mean through exponentially weighted noise. Its weights include the Gaussian importance correction as well as summed reward. It returns the updated mean’s first action. A smaller temperature sharpens the reward contribution to these weights; it does not change the exploration standard deviation. Changing the scene’s reward scale can change an appropriate temperature.

iCEM and MPPI operate in normalized [-1, 1] coordinates and map back to each channel’s bounds. Thus noise 0.5 means half the channel’s half-range, before clipping. Zero normalized input is the channel midpoint, which can differ from neutral input. In deadline mode, shooting uses a completed round’s action when available, otherwise its current mean. A partial batch’s attractive future is not automatically the action that gets executed.

These baselines produce future-state clouds and executed-world recordings but no search trees. The random baseline’s cloud is just its root state. Large shooting settings are checked against a 128 MiB sample-array limit; that is one array’s limit, not a bound on total process memory. Reduce walkers or horizon if allocation is rejected. Use Experiments, comparisons, and performance to measure quality, simulated work, and elapsed planning time on the same tasks and seeds.

Camera, manual control, and observation layers

Camera and layer controls change presentation without changing physics or resetting the run. Right-drag horizontally to look around the current view center, or vertically to change the tilt. The camera stays upright and orbits that center; it does not fly freely through the scene. Scroll to zoom, and left-drag outside Edit to bring another part of the environment into view without changing the zoom. In Edit, left-drag moves draft scene objects. Middle-drag or Alt-left-drag pans in every mode. Right-drag only changes the viewpoint.

In an ordinary planar scene, 2D / 3D snaps the camera to its preset overhead or angled view. In flight mode, Side / overhead switches between a side-on view (showing altitude against the horizontal direction) and a view of the physics plane from above. You can rotate from either preset. All these views show the same physical state; changing the angle adds no dimensions to the simulation. If you turn the flight plane almost edge-on, clicks and dragging on that plane are ignored: a small pointer movement could otherwise jump a long way across the scene. Right-drag away from that angle to select, edit, or pan again.

Follow agent follows the selected body, or the first controlled body when none is selected. Right-drag changes the angle around that body while following continues; panning disengages following so the camera stays where you put it. Your angle survives simulation updates, replay scrubbing, style and live reward changes, and window resizing. Loading or restarting a scene restores its initial view. Whole arena and Reset view also restore that initial preset, the arena center, and the default zoom. The framing uses the collision boundary’s bounding box with a 5% margin around the playable arena, recalculated for the viewport. Think of placing a snug rectangle around the actual fence, then widening it just enough to leave breathing room.

In Inspect, a click selects a body; select Planner decisions in the timeline to inspect exploration nodes instead. Scene edits support Undo/Redo and remain a draft until Apply and restart. Leaving a dirty editor offers Apply and restart, Discard, or Cancel.

Inspect / Edit / Drive describes how you interact with the world; Live / Replay describes which world you are looking at. Opening a saved run gives read-only replay. Use Create run from this frame before editing or driving that historical state.

Enter Drive to pause autonomous execution and open the selected vehicle’s controls in the inspector. With no controlled vehicle selected, driving uses the first controlled body. Selection is shared with camera follow and action guides. Click the world if a text or numeric field has focus: typing into settings should not steer a vehicle. The inspector shows the supported keys and pointer-operated buttons, together with actuator sliders.

Start driving starts the worker’s physics clock. It advances fixed scene.physics.dt frames independently of keyboard repeat. Releasing the keys returns the command to neutral; the vehicle can coast, fall, or continue moving under its existing forces. Neutral input is not a brake. Each timer callback advances at most five frames and discards excess catch-up backlog, so a slow tab can run slower than wall time rather than suddenly jumping ahead. Actual executed motion enters the recording.

Pause stops physics and clears held input. Losing focus, hiding the tab, opening a modal, or leaving Drive also pauses and clears commands. Selecting Inspect returns to paused controller operation, with stale plans discarded before the next run. While paused, Step physics frame or Apply action · 1 frame advances exactly one frame using the proposed actuator command.

Input	Channel mapping
W / S	Positive/negative thrust, throttle, or body-local force_x, clamped to bounds. A forward-only rocket cannot reverse its thrust channel; in flight, thrust must also supply the upward force needed to counter downward gravity.
A / D	Positive/negative torque or steering.
Q / E	Positive/negative body-local force_y.
Space	brake = 1 where a brake channel exists.
Drive controls → Apply action · 1 frame	Set named sliders in the inspector, then commit one physics frame while paused.

Channels outside the selected body receive neutral inputs during keyboard control. The live keyboard adapter recognizes the channel names listed above. Independent thruster_N channels currently require the Actuator channels sliders: the native channel descriptors supplied to the adapter include names and bounds, but not the thruster geometry needed to combine drive, strafe, and turning input. Unrecognized custom channels also use sliders until a keyboard mapping is added. Sliders expose the selected controlled body’s channels, bounds, and current proposed value, with increments of one two-hundredth of each channel range. In paused Drive, moving a slider changes the proposed command without advancing physics. During continuous driving, the new command takes effect on subsequent physics frames.

Layer/control	Default	Visible meaning
Animations (beside Visual style)	On; off when the system requests reduced motion and no explicit choice is saved	Enables cosmetic vehicle motion and effects across live views, replay, comparisons, and the workshop. Your explicit choice persists in this browser. Off freezes decorative motion and skips its updates. Static thrust and individual-jet cues, steering, reverse/brake lamps, and enabled action guides still follow current commands; native movement, cargo amounts, pickup visibility, tether connections, and diagnostics still update. Toggling does not reset the simulation or change recordings.
Action guides (beside Animations)	Off; explicit choice persists across Lab tabs	Shows signed command arrows and a numeric readout for the selected controlled body, falling back to the first controlled body. Percentages are relative to each channel’s configured action limits, not measured forces. Guides remain available with animations off.
Pause with animations on	Gentle idle motion continues	Stops simulation-dependent wheel motion and event progression while retaining small hover and engine motion. Turn Animations off for a still presentation of the paused world.
Workshop Play animation / Pause animation	Stopped initially	Advances or pauses decorative motion using the current actuator sliders; playback does not change commands or simulate physics. The global Animations switch must be on to play. Selecting Side or Top resets decorative motion while preserving slider values.
Workshop actuator sliders and Neutral / Max	Neutral commands	Shows the selected asset’s catalog actuator channels with their signed bounds. Neutral sets all channels to zero; Max sets each to its upper bound. Rocket variants expose vector thrust/torque or individual thrusters. Values persist across asset, style, and detail changes during the workshop session; static action cues update even with animations off.
Rollout paths	On	Recorded controlled-body paths, available for FMC and Wave Jump. Green is at/above mean recorded reward; rose indicates terminal state; violet indicates a tethered path; blue shows other alternatives. Terminal color takes precedence over tether color.
Future-state cloud	On	Controlled-body positions in the planner’s returned world batch, which may be at a partial horizon in deadline mode.
Tethers & formation	On	Physical tethers plus, for task: "tandem" with at least two controlled bodies, dashed links between every scored pair’s displayed centres. Link colors show pair quality; turning the layer off hides its links and legend.
Collision geometry	Off	Native hull outlines, useful when a decorative model differs from its collider.
Clean view / Show diagnostics	Diagnostics visible	Temporarily hides these four layers, then restores their previous visibility. Checkpoint indicators and the separate Physics inspector remain independent.

Formation links use a continuous, fixed color scale in both visual styles: rose/red at 0, amber at 0.5, and green at 1, the perfect pair distance. The matching Pair quality: 0 — 0.5 — 1 · Perfect legend appears beneath Tethers & formation when formation links are enabled and the scene has at least two controlled bodies with task: "tandem". Every unordered pair appears once, with endpoints following the displayed centres live, paused, and in replay. Colors use the configured pair targets and their scalar fallback; the scores multiply to give total formation quality. They remain visible at formation reward weight zero. Physical tethers retain their existing appearance and behavior.

In Tandem flight, the shared active checkpoint has an amber ring and fill, while inactive checkpoints stay subdued. A label such as Checkpoint 2 · 1/3 crossed identifies the checkpoint and how many controlled vehicles have already crossed it. These indicators follow the displayed world in live, paused, and replay views.

Each registered crossing flashes that checkpoint green, fading over 600 ms. When the final vehicle arrives, the amber highlight immediately moves to the next checkpoint while the completed checkpoint’s green flash fades. The next physics frame then begins using the newly unlocked checkpoint, as described above.

Forward replay playback shows crossing flashes. Seeking or scrubbing to a replay frame, or returning to live view, updates the checkpoint indicators without inventing crossing flashes for the jump. Turning Animations off or requesting reduced motion in the operating system suppresses these transient flashes; the active highlight and crossing-count label still update. Checkpoint indicators remain visible when Tethers & formation is off and in Clean view.

The renderer samples large trees to draw roughly at most 50,000 path segments. This drawing limit does not prune the native record. Hiding paths also does not disable recording. For that, change Tree in the Planner decisions timeline and commit the change with Apply and restart.

Read the telemetry and choose recording detail

Task progress, simulation time, and planning progress remain near the world. Open the inspector’s Diagnostics for population, risk, memory, and performance measurements. Selected decision shows recorded action, path reward, outcome, and reward weights when available. Missing values read Not recorded; individual reward contributions are not inferred from the weights or recomputed for old frames. The timeline separates World motion from Planner decisions, whose cursors count different things: physics frames and controller decisions.

For Wave Jump, the selected path reward describes the chosen final walker and trajectory progress describes execution of the full path, shared prefix, or one-action fallback. The status identifies shared-prefix execution or fallback and shows search depth. These are separate from search progress: completing the search starts the journey.

Three measurements answer different questions: simulation time measures what the world has done, planning progress measures search work, and risk probes examine short random continuations of the chosen action. A cloud with many dead worlds does not by itself measure the risk of the action ultimately selected.

Readout	Exact interpretation
SIMULATION / TICK	Executed or replayed world tick times physics.dt; not wall-clock runtime.
DEAD RATIO	FMC/Wave Jump final Wave terminal fraction. Shooting reports terminal rows divided by allocated Walkers; iCEM’s inactive capacity can dilute this fraction. Random has no searched population, so its zero is not evidence of safe exploration.
SELECTED-ACTION RISK	Fraction terminal in 16 separate continuations: chosen action for F frames, then independently sampled uniform joint actions held for 2F frames, where F = Action frames. Display includes sample count and total horizon.
EVAPORATED / CLONED	Nodes removed by the latest FMC pruning pass, and fraction cloned in the latest Wave iteration. Wave Jump shares these operations; shooting and random controllers do not.
AI BUDGET USED	FMC/Wave Jump iterations / Horizon; shooting completed depth advances / (Horizon × Search rounds); random reports 100%. Capped at 100% in the HUD. This is search progress, not CPU utilization.
ITERATIONS / MS	FMC/Wave Jump Wave iterations or shooting depth advances, with measured planner elapsed milliseconds. Live elapsed time includes worker yields and risk evaluation.
SCORE	Scene-defined task readout: for example deliveries, pickups, gates, or completed laps. See each task in Scenes, agents, and the editor.
BYTES / WORLD	Unpadded mutable world-state size. Excludes shared scene data, serialization header, planning arrays, and graphics.
Footer	Body count, joint action dimension, and missed real-time action deadlines.

The risk statistic is a short empirical test under random continuations, not a calibrated failure probability under future planner decisions or a safety guarantee. It has resolution 1/16 = 6.25%. In Wave mode it is unevaluated. Historical decisions display their recorded diagnostics when available, not a fresh test of the currently visible replay frame.

Physics inspector shows velocity, external force, nearby contact normals, and signed spring/damper tether force. Its text summarizes the selected or first controlled body: velocity in m/s, angular velocity in rad/s, force in newtons, nearby contact count, and peak absolute tether force. Cyan arrows use velocity times 0.25 s; amber and magenta use force times 0.05 m/N. Red normals indicate current nearby geometry, not reconstructed past collision impulses.

The live performance line reports FPS, CPU render submission time, draw calls, triangles, authoritative-engine world frames per measured stepping second, and tracked native buffer bytes. Submission time is not GPU completion time, and tracked buffers exclude some scratch/tree/allocator overhead. In Experiments, control effort means integrated squared channel input, not mechanical energy; planning world-frame counts and throughput probes provide separate work measurements.

Recording control	Operation
Tree → Pruned (recording = 1, default)	Remove orphan leaves while protecting current walkers, elites, and their ancestors.
Tree → Full (recording = 2)	Retain all FMC/Wave Jump search nodes; useful for examining alternatives, with higher memory use.
Tree → Off (recording = 0)	Disable visible search-tree recording. Wave Jump still uses at least pruned recording internally to reconstruct its trajectory. Executed-world recording continues.
Keep all decisions	For memory-backed recordings, replace the usual latest-32-decisions window with retention up to the 64 MiB tree payload budget. Capacity stops control with an export prompt; it does not silently discard the oldest retained decision.

Without Keep all decisions, older trees leave the in-memory window when either 32 decisions or the tree budget is exceeded. A single oversized tree is still an error. Device-backed recordings store trees separately and keep only a bounded recent window resident. World motion has its own storage budget and remains available even when a controller provides no tree. Follow Record, replay, and continue experiments for world playback, branch reconstruction, snapshots, planner checkpoints, portable files, device storage, and event markers.