THE COMPOSITION AXIS — PRESENTER SCRIPT Controller-Guided, Composition-Agnostic Multi-Agent RL for Drone Swarms in a Physically Grounded Digital Twin Zahid Hussain · Texas A&M University How to use this --------------- - One block per section of the site, in scroll order, from basic to advanced. - [SCROLL] says where to be. [DO] is an action on screen. [NOTE] is for you, not the audience. Everything else is what to say. - Timings are a guide for a full walkthrough of about 40–45 minutes. For a 15-minute version, do only the sections marked (*). - Every number below is from the current draft. Learned-policy results are PENDING in the paper — never imply otherwise. Only the hand-written ("scripted") and random reference values are final. - The paper is single-author; the script says "we" by academic habit — swap in "I" if you prefer. - The menu (top right, "Sections") jumps to any section if you need to skip. ===================================================================== HERO (*) ~1 min ===================================================================== [SCROLL] Top of the page. Let the 3D swarm fly for a few seconds. "This is The Composition Axis. The goal is one AI pilot that can fly a team of drones of any size, split between two teams in any way — on simulated drones that behave like real multirotors, in games where cheating doesn't pay. Behind me is an illustration of the idea: the swarm cycles through line-ups — five versus five, four versus six, three versus three — with unused slots parked, and one brain is meant to fly all of them. It's an animation, not a trained policy; the learned results are still pending. Three numbers to remember: seventeen team line-ups, all practised by one policy; five games; one simulator shared by training, evaluation, replays and a live browser twin." [NOTE] The "now flying … same AI brain" chip describes the goal, not a result. ===================================================================== START HERE (*) ~2 min ===================================================================== [SCROLL] "Everything you need to follow this page". "If you're new to drones or AI, six ideas are enough. One: reinforcement learning — the AI tries an action, sees what happened, gets a score, and over millions of tries does more of what scored well. Two: the AI's brain is a policy — a neural network that takes what one drone senses and outputs what it should do. Every drone runs its own copy of the same brain. Three: a real quadrotor can't just move sideways. It has to lean, leaning takes time, and wind, drag and a draining battery all get in the way. Most AI studies ignore this and simulate idealised dots. Four: teams. In cooperative games every drone is on one side — fly in formation, keep watch. In competitive games two teams play each other — chase, capture-the-flag, hold zones. Five: all practice happens in a simulator — a digital twin — because real drones crash and cost money. Six: for every game, engineers also wrote a hand-written pilot — a scripted controller. It's decent, not brilliant. The AI first learns from it, then has to beat it. Beating it is how success is measured." [DO] Point to "the page, as a map". "The page goes in five parts: the problem, the world, the game from a drone's eyes, how the AI is built and trained, and how results are judged. Underlined words have definitions on hover." ===================================================================== 01 — THE PROBLEM (*) ~2 min ===================================================================== [SCROLL] The three problem cards. [DO] Click each card. "Three shortcuts keep AI drone teams away from real deployment. Team size: most policies only work at the exact team size they were trained for — fixed input slots for drone #1, drone #2. Change the roster and you retrain. The drones: most are trained on idealised dots that accelerate sideways as hard as they climb and brake downward faster than gravity. Skills learned on dots don't carry over. Loopholes: AIs are very good at scoring without doing the job — crashing on purpose because the crash penalty is smaller than staying alive, hovering where a rule forgot to look, or being excellent in one role and useless in the other while the average looks fine. Each shortcut is known on its own. A system that removes all three at once, across cooperative and competitive games, hadn't been described." [SCROLL] The box under the cards. "The central idea is the composition axis: treat 'how many drones, and how are they split between the teams' as something to practise across. With 6 to 10 drones and each team holding 30 to 70% of them, there are exactly seventeen line-ups — few enough to practise every one and keep score on each." ===================================================================== WHAT THE PAPER CONTRIBUTES (*) ~1 min ===================================================================== [SCROLL] The five contribution cards. "Five pieces. One: a realistic, checkable drone simulator. Two: CAST, one brain for any team size, with PCR, a practice scheduler over line-ups, and a league organised by line-up. Three: CASH, one brain for different kinds of drone. Four: CGPO, learning from the hand-written pilot. Five: scoring rules that mean what they say. We claim no novelty for the set-transformer itself. The contribution is the composition axis as a training resource, the controller as a training participant, and the physical and procedural guarantees that make results on this axis meaningful." ===================================================================== 02 — THE VEHICLE ~3 min ===================================================================== [SCROLL] "The physics, in one line". "The AI never touches the motors. It asks: accelerate this way, this hard — the same instruction a real autopilot accepts. The simulator then works out what a real drone would actually do. The drone can lean at most 40 degrees, its rotors push between 0.1 and 2.2 times gravity, and it catches up to a new lean over 0.12 seconds and a new push over 0.04 seconds, hearing the command one physics step late. Then air resistance — linear rotor drag and quadratic body drag, measured against the moving air, so wind matters — and the downdraft from drones above." [DO] Figure (a): drag the requested acceleration; toggle the three effects off and on. "Here's the cost of realism. Both are told 'full speed ahead'. The idealised dot obeys instantly; the real drone has to lean first, and can only lean so far. Toggle each effect off to see what it costs." [DO] Figure (b): drag the circle to ask for hard sideways plus hard downward. "This is which moves are physically possible. The dashed box is everything the AI is allowed to ask for; the green fan is what the airframe can deliver. Ask for hard sideways and hard downward at once and it's cut back — a drone can't brake downward at full sideways acceleration. Holding height, it manages about 8.2 metres per second squared sideways, and it can speed up downward by at most about 8.8 — a little less than falling." [DO] Figure (c): drag "how far below". "And the downdraft: a drone directly under another is pushed down by up to 0.6 g, fading to nothing four metres down. A drone flying underneath a teammate has to spend thrust just to hold altitude." [SCROLL] "Varied for every drone, every round". "Wind is a steady breeze up to 2 metres per second from a random direction plus 1 metre per second of gusts. The battery sag costs up to 20% of thrust. And each drone's strength, lags and drag are randomised every round — plus or minus 10 and 30 percent — so the AI can't rely on one exact airframe. That's the standard recipe for skills that survive on real hardware." ===================================================================== 03 — THE SHIELD (*) ~2 min ===================================================================== [SCROLL] The safety bubble. [DO] Drag approach speed up, then braking power. "Penalties alone don't keep drones apart — an AI will accept a small penalty if it wins more elsewhere. So a safety guard sits between every pilot, AI or hand-written, and the drone. Around each obstacle it keeps a bubble: a 4 metre buffer, plus the distance needed to brake from your current approach speed, plus room for a 0.4-second reaction delay. Speed up and the bubble grows. Inside it, any push toward the obstacle is removed." [SCROLL] The four cards. "The six arena faces get the same treatment — 2 metres of margin, 3 above the floor. Two teammates are kept at least 2 metres apart plus their stopping distance. Opponents are deliberately not protected, so chasing, tagging and blocking stay possible." "A subtle but important detail: the guard changes what the drone does, not what the AI asked for. The AI is still scored on its own request, so the learning maths stays correct — and it has to discover for itself which requests get overruled. Training under the filter, not just deploying with it, follows CBF-RL." ===================================================================== 04 — THE ARENA ~1 min ===================================================================== [SCROLL] The obstacle layout. [DO] Press "new round, new layout" a few times. "If the obstacles never moved, the AI could memorise the map. So every round, 3 to 8 pillars and boxes are dropped inside the central 40 by 40 metres, at least 6 metres apart and clear of the formation path, bases, pads and zones. Avoidance has to be read from the drone's sensors, every time." [SCROLL] "What a drone can actually sense" and the arena card. "The arena is 60 by 60 by 30 metres — about half a football pitch, three storeys high. The AI decides every 0.1 seconds; physics runs five times in between. Everything random comes from a seed, so any round can be replayed exactly." ===================================================================== 05 — THE DECISION PROBLEM (*) ~2 min ===================================================================== [SCROLL] "The game, in seven parts". "Formally this is a partially observable Markov game — seven parts. Who plays: every drone in this round's line-up, with room always kept for ten. What's really going on: the full truth — only the simulator has it. What one drone perceives: a list — about itself, other drones, objects. What it can do: three numbers between −1 and 1. How the world responds: safety guard, then realistic physics, every 0.1 seconds. How it's scored. And how much the future counts: gamma = 0.99." [SCROLL] "Who sees what". [DO] Toggle "During The Game" / "During Practice". "During the game each drone decides on its own, from what it senses and what teammates radio in. During practice, a coach — the critic — sees the whole truth to judge how well the game is going. It's discarded when the AI flies." [SCROLL] "How each drone is scored, every moment". "Points for the task, small hint points for progress — built as potential-based shaping, so hints can never change what the best strategy is — and penalties for bumping into things, crashing and flying close to walls." ===================================================================== 06 — WHAT A DRONE SEES (*) ~4 min ===================================================================== [SCROLL] The draggable 5-v-4 scene. [DO] Drag the ringed drone; click a radio link to cut it. "A drone never sees other drones directly. It spots them only within 15 metres, only if no obstacle is in the way, a tenth of a second late and a little off — the noise grows with distance. Teammates share what they spot over a 20-metre radio that drops 5% of messages, and opponents never relay. From those sightings each drone keeps a running best guess of where every other drone is — a track. Watch the counters: drones I see, drones known by radio, and drones that are simply switched off in my input. Cut a radio link and a drone vanishes from my picture." [SCROLL] The token cards on the right. "What the network receives is three kinds of item. About itself: 43 numbers — its motion, 10 obstacle beams out to 12 metres, 6 wall distances, team and clock, its top speed and acceleration, the hand-written pilot's suggestion, and 8 task numbers. About other drones: 9 slots, nearest first, each with where I think it is, how it's moving, teammate or not, and how old the guess is. About objects: up to 8 — zones, flags, pads, points of interest." [SCROLL] The state table. "This is one drone's complete input, field by field, for this exact scenario: format, size, and the value it receives right now — 222 numbers in all, the same layout for every drone, every line-up and every task." [NOTE] Two rows in this table reflect an earlier code layout and differ from the current draft. Don't read them out as written. Say instead: - "Opponents alive" — the paper now says this is the scoreboard's count, not the truth: a crash nobody saw isn't announced, so the number can be wrong. - Pursuit chaser task numbers — the paper now says they are computed from the drone's own tracks, like everything else about other drones. [SCROLL] "How a track is kept — and lost". [DO] Drag time; toggle "a teammate is watching". "Here a chaser follows a runner that slips behind a pillar. The hollow circle is the chaser's guess. While the runner is visible, an alpha-beta tracker nudges the guess halfway toward each sighting. Once it's hidden, the guess coasts at the last speed — and drifts off when the runner turns. After 3 seconds with no sighting it's dropped. A sighting after more than two steps of drifting, or outside the gate, restarts the track instead of dragging it. Turn on the teammate and the guess stays fresh — which is why obstacles become tactical, and why the radio has to share enemy sightings, not just teammates' positions." [SCROLL] "What the referee tells everyone". "Game events don't reach the drones by magic either. Like a real match, a referee announces tags, stuns, landings and flag events. A carried flag is known from a beacon once a second, with 2 metres of noise. Zone ownership is broadcast once a second, but never how many drones are inside. The hand-written pilots work from exactly the same information — only the coach, during practice, sees the truth." ===================================================================== 07 — WHAT A DRONE DOES ~2 min ===================================================================== [SCROLL] "From the AI's decision to actual movement". [DO] Move the three sliders; tick the obstacle box. "Every 0.1 seconds the AI gives each drone three numbers: how hard to accelerate east–west, north–south and up–down — the arena's fixed directions, not the drone's heading. Multiply by that drone's maximum acceleration and you have the request. Then the safety guard removes anything dangerous, and the airframe cuts back what it can't physically deliver. The four panels show what's left at each stage." [SCROLL] The action table, then "How the AI picks a number". [DO] Toggle "Adds Corrections". "The AI actually outputs a best guess and a spread for each axis, and picks around the best guess — that randomness is how it keeps exploring. In the 'adds corrections' version the action is the pilot's command plus a quarter of the AI's, and the allowed randomness is tighter. Why 'accelerate' and not 'spin motor three faster'? Real drones already have autopilots that speak this language, and it lets the AI and the hand-written pilot share one action space — so the AI can copy the pilot, be compared with it, or add corrections to it." ===================================================================== 08 — THE COMPOSITION AXIS (*) ~3 min ===================================================================== [SCROLL] The practice scheduler grid. [DO] Press "run training". "This is the heart of it. Every box is a line-up — 3 versus 5, 5 versus 5 and so on — from 6 to 10 drones. The bar is how often the AI currently wins it; the darker the box, the more often it gets practised. The scheduler, Prioritized Composition Replay, is like a good coach: it avoids drills the AI has mastered and drills that are hopeless for now, and focuses on the ones it wins about half the time. The score is p times one-minus-p — zero if it always wins or always loses, highest at fifty-fifty. Line-ups are ranked by that score and picked mostly by rank — but 30% of the time by how long since they were last played, so nothing is forgotten." [SCROLL] "The line-ups" card. "Totals 6 to 10, each team 30 to 70% — in the chase game chasers hold 50 to 70%, so they're never outnumbered. That's 17 team-vs-team line-ups and 5 sizes for the one-team games, with room always kept for 10 drones: unused slots are parked — invisible, unable to collide, and ignored by learning. Other methods practise small teams first and then stop; this scheduler never stops covering every line-up." ===================================================================== 09 — CAST ~2 min ===================================================================== [SCROLL] The actor diagram. [DO] Toggle "Pilot" / "Coach". "The network, CAST, reads each drone's view as a list — itself, other drones, objects — and a transformer, the same kind of network behind modern chatbots, decides which items matter: 128 wide, 4 heads, 2 layers. Nothing in it says 'drone #3', so a 3-drone team and a 10-drone team look the same to it — just shorter or longer lists. The coach is built in the style of the Multi-Agent Transformer: it reads every drone at once, from the true positions rather than guesses, and gives each drone its own value. It exists only during practice." ===================================================================== 10 — LEAGUE AND CAPABILITY ~2 min ===================================================================== [SCROLL] The match-up grid. [DO] Press "pick a match and play it" a few times. "In team games the AI practises against saved copies of itself, plus the hand-written pilot. Every 50 updates a copy goes into a pool of 8. In half of all practice games one team is played by a pool member. Each match is picked as a pair — which opponent, and which line-up — and the pairs the AI still loses get picked most. Watch the picked match improve and fade. The hand-written pilot never leaves the pool: it's picked as long as the AI loses to it. Elo ratings are kept overall and for every line-up." [SCROLL] "CASH". "Chasers and runners have different limits — in the chase game chasers are faster, runners more agile. Instead of separate brains, each drone passes its top speed and acceleration into a small helper network — a hypernetwork — that adjusts the final decision layer. One brain, different drones, and it can even fly limits it never practised with." ===================================================================== 11 — CGPO (*) ~3 min ===================================================================== [SCROLL] The four "ways" cards. "Starting from scratch wastes the hand-written pilot twice: the AI has to rediscover basic flying, and practising only against itself makes it good at beating itself rather than the pilot it'll be judged against. So the pilot joins training in four ways. Copy first: the pilot flies 200,000 steps and the AI learns to reproduce its moves, over 20 passes. A gentle pull back: straying from the pilot costs a little — the pull fades over the first 30% of training to 5% and stays there, like a light tether that's never cut. Let the pilot fly the start: in half the rounds at first, the pilot flies the opening — up to 100 steps — and that fades. And keep it as a sparring partner in the league." [SCROLL] "How the pilot's help fades". [DO] Drag training from 0 to 100%. [NOTE] The paper gives the start values and says the guide episodes are annealed; the site's fade-to-zero-at-30% curve is its own simplest reading. Don't state an end point for it. [SCROLL] "Two versions". [DO] Toggle "Flies alone" / "Adds corrections". "Every task is trained both ways. Flies alone: the AI learns with the pilot's help but flies on its own — the only version that can do better than the pilot where the pilot is weak. Adds corrections: the action is the pilot's command plus a quarter of the AI's, starting from exactly the pilot — the safer choice where the pilot is strong everywhere." ===================================================================== 12 — THE TRUST GATE ~3 min ===================================================================== [SCROLL] The trust gate equations and "one moment". [DO] Drag "pilot's edge" across zero, then "coach still learning". "The pull toward the pilot is the same everywhere, but the pilot isn't equally good everywhere: the formation pilot is excellent holding still and clumsy when things move fast; the chase pilot may be strong on one side and weak on the other. The optional trust gate lets the coach learn one more skill: scoring a specific move in a specific moment. At every moment it compares the pilot's move with the AI's. If the pilot's move scores as well or better, keep pulling toward it. If the AI has found something better, stop pulling — with no floor. Early on, while the coach is still learning, a fixed schedule does the pulling." [SCROLL] "Every moment at once". [DO] Drag training updates to 10. "In a quick test run, within ten training updates the coach was already confident in about one moment in five each way — trusting the pilot in some, the AI in others. The bars here are illustrative; that one-in-five figure is the paper's. These are exactly the numbers it will report when it tests the gate on formation and the chase game, in both versions. The cost is one extra coach calculation per batch." ===================================================================== 13 — THE FULL ITERATION ~2 min ===================================================================== [SCROLL] The loop. [DO] Press "auto-play", or click steps 0–7. "All of it in one cycle — the paper's Algorithm 1. Once, at the start: copy the pilot. Then, thousands of times: pick an opponent and a line-up; set up the rounds — which line-ups, how hard, how often the pilot flies the start; play under the realistic drone and safety guard; judge each move with agent-terminal credit; improve with PPO plus the pull toward the pilot; update the schedules; and every so often, test at full difficulty on the standard line-up and keep the best. Three things to notice: the pilot helps in several places but is gone when the AI flies; two practice schedules run at once — who plays, and how hard; and the AI is always judged on the real, full-difficulty task, on each side of the game." [SCROLL] The training settings. "The settings: MAPPO, 64 games side by side, 128 steps each, 5 passes in 32 chunks, gamma 0.99, lambda 0.95, learning rate 3 times 10 to the minus 4 decaying to 10%. Each run is 30 million drone decisions — parked slots don't count — and every setting is trained three times. Nothing is switched off to make learning easier." ===================================================================== 14 — AN HONEST SIGNAL (*) ~4 min ===================================================================== [SCROLL] Rule 1. [DO] Toggle "usual maths" / "the fix"; drag the guessed future. "Given any loophole, an AI will find it. Three rules close the three most common ones. Rule one: leaving the game must never look like a good deal. When a drone crashes, is tagged or lands mid-round, the standard maths keeps estimating its future from a state the coach was never trained on — a blank slot or a runner parked on its pad — and that guess drifts toward 'average'. So leaving early looks like a free average score, and the AI learns to crash. Here crashing looks worth more than staying alive. The fix is one switch in the formula — e for 'still in the episode': once a drone is out, no imaginary future is added. A stunned capture-the-flag drone hasn't left: its steps stay in the chain with zero reward, but it isn't trained on actions it couldn't take." [SCROLL] Rule 2. [DO] Drag "points per zone held". "Rule two: winning must clearly pay more than losing, and nothing can be farmed. Small hint points that each look sensible can add up to more than the goal — you can't see it by reading the rules, you have to measure it. Play a strategy that wins and one that deliberately loses — for zone control, the hand-written pilot against a team that camps on one zone — and make sure the winner earns clearly more than the normal noise. That's how the zone weight was set to 0.3, exactly the rate the real score moves. Hints only count toward targets that stay put, tradeable bonuses are zero-sum, and anything that removes a drone from the game pays exactly what it gives up." [SCROLL] Rule 3. [DO] Click "balanced", "collapsed", "lopsided". "Rule three: judge the AI by its weakest side. Averaging both sides gives the same score to an AI that's good on both and one that's perfect on one side and hopeless on the other. So the score is the worse side's win rate, plus a tie-breaker worth less than one test game — 0.9 over n times tanh of twice the margin. It only orders AIs with the same weaker-side win rate. It can't rescue a side that never wins." ===================================================================== 15 — THE SECOND CURRICULUM ~2 min ===================================================================== [SCROLL] The difficulty dial. [DO] Drag difficulty from 0 to 2. "A brand-new AI chaser, trying to get within 1.5 metres of a runner in a 60-metre arena, basically never succeeds — and with no successes, it has nothing to learn from. So the chase game has a difficulty dial. From 0 to 1, runners go from half speed to full speed. From 1 to 2, the tagging distance shrinks from 4 metres to 1.5. Tests always run at 2, so every reported score is on the real game." [SCROLL] "What advances it". "The dial goes up by 0.1 whenever the chasers have won more than 70% of their last 40 rounds, and it never goes back down. Why not use the overall win rate? When the AI plays itself one copy always wins and one loses, so the overall rate sits near 50% no matter what. It says nothing about the chasers. So the dial watches the struggling side directly. The curve here is illustrative; the rule is the paper's." ===================================================================== 16 — THE FIVE GAMES (*) ~3 min ===================================================================== [SCROLL] The game tabs. [DO] Click through all five. "Two cooperative games and three competitive ones, all on the same simulator and sensors, any line-up, each reporting a simple win or loss for the practice scheduler. Formation: six drones, 40 seconds, fly as a shape and hold it, in four stages — hold still, move the whole shape, spin it, change it into a different shape. Scored by the average distance to your spot over the second half, and a crashed drone counts as at least 10 metres off, so dying never helps. Coverage: eight drones keep eight points watched. The hint pays exactly for staleness removed and charges staleness accrued, so letting points go stale and revisiting them earns nothing. Pursuit: three chasers against three runners. Chasers are faster — 10 against 8 metres per second — but less agile. Chasers tag within 1.5 metres; runners win by landing on one of two pads, and the air above a pad is contested, so landing is a real manoeuvre. Landing pays what the runner gives up by stopping, so landing never loses to loitering. Capture-the-flag: five versus five, first to three captures, 60 seconds, tagging in your own half, 3-second respawns. Pickups need you near the flag's height, and pickup and return bonuses are zero-sum. Zone control: five versus five, three zones of 6 metres, scored per second by the team with a strict majority inside. Ownership reaches the drones only through the once-a-second broadcast." [NOTE] If asked about the hand-written pilots: they act per radio-connected group, on that group's tracks, through the same safety guard — PID-to-slot, greedy stalest points, attacker/defender splits, force concentration, and a pursuit controller that aims at the closed-form intercept point. ===================================================================== 17 — EVALUATION PROTOCOL (*) ~3 min ===================================================================== [SCROLL] The seven questions. "The protocol was fixed before any results exist. Seven questions: does one AI do equally well on every line-up; does it work beyond the practised sizes — trained on 4 to 8, tested at 10; is one brain told each drone's limits as good as separate brains; where does each version beat the pilot, and on which side; does the trust gate help; does easy-first make the chaser learnable; and what does each piece add." [SCROLL] "Choosing the best saved AI" and "How final scores are reported". "A snapshot is tested every 40 updates on 16 repeatable games at full difficulty and the standard line-up; the one kept has the best average over three tests in a row. Final scores: 20 fresh games per seed, seeds never used in training or selection, three seeds, reported as mean plus-or-minus spread — and team games are reported per side, never just averaged." [SCROLL] "The cheating check". "Before any score counts, the AI is checked for loophole behaviour: hugging the ceiling or floor, bumping, crashing, maxed-out commands, just copying the pilot, and task-specific signs — landing altitudes and late loitering in pursuit, pickups that never become captures, camping on one zone. The sharpest sign: do high-scoring rounds actually do the task? If points and real performance stop moving together, the AI is earning from hints, not the goal. The same signs are logged every training episode, so an exploit shows up while it's emerging." [SCROLL] "The implementation audit". "A behaviour check can't catch a learner that's wrong in ways that still train. So before the reported runs, the observation pipeline, the learner and the simulator were reviewed line by line, and every confirmed defect was fixed with a regression test. Five examples: a mask read from the wrong feature that hid nearly every visible drone from the network; a coach that saw different inputs while playing and while learning; samples routed to the wrong team's model; a chaser input that ignored obstacles; and track velocities that exploded after re-acquisition. The training runs were restarted from scratch after the audit." [SCROLL] The planned ablations. "And the planned remove-one-piece experiments: no hand-written pilot, no speed-and-agility input, no practice scheduler, no randomness — plus transfer grids per team game, and a capture-the-flag test trained on 4 to 8 drones and played at 5 v 5, 4 v 6 and 3 v 7." ===================================================================== 18 — REFERENCE VALUES (*) ~3 min ===================================================================== [SCROLL] The status box and the two charts. "Status, plainly: the AI training runs — five games, both versions, three repeats each, 30 million decisions per run — are in progress after the audit. Every AI score in the paper is marked pending, and so is every AI score here. What's final are the scores of the hand-written pilot and of random flying, measured in exactly the simulator the AI trains and is tested in, 20 rounds each on held-out seeds. Those are the bars to clear." "Formation: under gusting wind, the shield and new obstacles every round, the hand-written pilot holds a still six-drone shape to within 0.65 metres and a spinning one to 1.05. But it falls behind when the shape moves — 1.47 metres — or changes — 1.64 — because a rule that chases where the spot is now, without anticipating, always lags a moving target. That's exactly where a learning AI has room to do better. Random flying leaves the formation 32 to 35 metres off. Coverage: 90% of points kept watched with a mean staleness of 1.3 seconds, against 22% for random. The chase is where imperfect sight hits hardest: working from tracks, the hand-written chasers win only 0.35 of games and 62% of runners reach a pad — a runner behind a pillar is lost unless another chaser sees it, and a chaser that sees nobody can only guard the pads. Capture-the-flag — red wins 0.55 — and zone control — 0.50 — are close to even, so an AI that wins more than half its games on both sides has genuinely beaten the pilot, not exploited a lopsided game. And no hand-written pilot ever crashed or touched a wall under the shield." [SCROLL] "What the paper itself says are the limits". "The limits, in the paper's own words: single GPU, single machine; transfer tested within the trained range and one step beyond; perception without a limited field of view, missed detections or false alarms, and a referee assumed reliable; a reduced drone model — no full rigid body, no ground effect — so hardware would need re-identified lags, drag and downwash and a re-derived shield; the corrections version inherits its pilot's flaws; the benchmark is a strong, fixed bar, not the best achievable play; the cheating check only finds the cheats it knows to look for; and the AI results are pending." ===================================================================== GLOSSARY and CLOSING (*) ~1 min ===================================================================== [SCROLL] The glossary. "Every term on the page, in plain words, in one place. To close: this system treats team composition as a first-class training and evaluation variable for drone swarms, on vehicles that fly like multirotors, under rules that don't reward the obvious exploits. One composition-agnostic brain with a learnability curriculum and a line-up-indexed league carries one parameter set across rosters; capability conditioning carries it across vehicles; and controller-guided optimisation turns the hand-written pilot every benchmark already has into a training participant. And three procedural rules matter as much as the recipe: an agent's own death ends its advantage recursion, winning must out-earn losing with nothing to farm, and team-game checkpoints are selected and reported per side after a behavioural audit. Thank you." ===================================================================== LIKELY QUESTIONS — short answers from the paper ===================================================================== Q: Where are the AI results? A: Pending. The full set of runs was restarted from scratch after the implementation audit and is in progress under the final environment. Only the scripted and random reference values are final, and they're measured in exactly the environment the policies train in. Q: Why not train on point masses like everyone else? A: Point masses can accelerate sideways as hard as they climb and brake downward faster than gravity. Skills learned that way don't carry over. Every command here goes through thrust and tilt limits, lags, drag, gusts, downwash, battery sag and latency, randomised per episode. Q: Why is the chase so hard for the scripted pursuers (0.35)? A: Perception. They act on tracks, not truth: an evader behind a pillar is lost unless another pursuer sees it, and a pursuer that sees no evader can only guard the pads. Q: What stops the AI from just crashing to avoid penalties? A: Agent-terminal credit: once a drone is out of its episode, no bootstrapped future is added, and a crash also charges the discounted stream of reward the drone would have collected. Q: Doesn't the safety shield change what the AI learns? A: It changes what the drone does, not the policy's output or its likelihoods; the action cost is charged on what's flown. The AI has to learn which requests get overruled. Training under the filter follows CBF-RL. Q: Why practise across line-ups at all — isn't one team size enough? A: Rosters change between missions and engagements are often asymmetric; a policy bound to one size and split has to be retrained when either changes. The axis is small enough to enumerate exactly — 17 two-team compositions — so the curriculum keeps exact per-composition statistics, and evaluation covers the whole axis, including splits the policy rarely samples. Q: How is this different from kickstarting or residual RL? A: Those reuse a controller, usually single-agent. CGPO puts the controller in four channels at once — initialiser, anchor, guide episodes and permanent league member — inside a multi-agent league, in two arms, with an optional critic-side trust gate. Q: Can the trust gate go wrong? A: While the Q branch is untrained a fixed schedule carries the anchor, and guide episodes keep a small permanent probability so Q stays calibrated on the controller. It's evaluated as a separate experiment on formation and pursuit, with its trust statistics reported. Q: How do you know the AI isn't cheating? A: A behavioural audit on every checkpoint — altitude, contacts, crashes, saturation, dependence on the pilot's action, return-versus-metric correlation and task counters — logged every training episode too. The paper is explicit that an audit only finds exploits it knows how to look for. Q: Would this transfer to real drones? A: The model is reduced — thrust direction with first-order dynamics, no ground effect. A hardware deployment would re-identify lags, drag and downwash and re-derive the shield envelope from the measured closed-loop response.