Ten lessons for an FTC team, from 3,300 simulated matches

We are the students and mentors of the three FTC teams at NCSSM: 5064, 8569, and 22377. This post is for a team that is deciding what to build, what to practice, and how to plan a match. We simulated BIOBUZZ, the 2026-2027 FIRST Tech Challenge game, about 3,300 times and changed one thing at a time. The game details are BIOBUZZ, but most of the lessons carry to any season in which robots collect, launch, and race a clock.

The Match Lab page shows the same results as charts, and it estimates the score of a robot that you configure.

Each lesson starts with the advice. The numbers follow, and then something to try at practice. Read the ranking as "where to look first", not as a prediction for your robot. The section What a simulator can't tell you lists what the numbers leave out.

How to read the numbers

• Points are per alliance, per match. A typical alliance in these matches scores about 410 points.
• One match varies by about 20 points. Most rows are 12 matches, so a difference under 25 points is unproven. We say so where it applies.
• The baseline robot is 38 cm15 in. square, 10 kg22 lb, and geared for 600 rpm at the wheel, with a rear-facing shooter that releases two elements at a time and a 95% reliable intake.

BIOBUZZ in one paragraph: robots collect POLLEN and NECTAR and launch them into the raised CELL of their HIVE. About 8 POLLEN tip the HIVE for 20 points, which dumps that CELL onto the floor and raises the other one. FLOWERS are a slower second objective. PARK is 5 points at the end.

The ten lessons

1. Make your launcher repeatable before anything else

Nothing else comes close. The baseline launcher varies by 0.08 m/s3.1 in./s in speed, 1° in elevation, and 1.2° in yaw. Three times that spread costs a third of the score.

A miss isn't one lost point. It delays a 20-point TIP, and the element has to be collected again. AUTO suffers most, because an open-loop AUTO can't correct: AUTO points fall from 88 to 42 at 3 times the spread.

Stop before you shoot, too. Letting the robot launch while it still moved at 0.35 m/s14 in./s cost about 18 points and saved no time.

At practice: launch 20 elements from one spot and measure the group. Then chase the causes: flywheel speed control, how each element seats, battery voltage, and worn wheels.

2. Find your scoring window, and launch from the middle of it

The window is a distance, not an angle. The baseline shot scores from any standoff between about 1.16 m46 in. and 1.58 m62 in. along the line of the shot. Launching from either edge of that window costs 96 to 98 points. A robot never stops exactly on its spot, and at the edge of the window a few centimeters off is a miss.

Angled launches are fine. In every match here, the second robot of each alliance launches from a spot 0.62 m24 in. to the side, about 40° off the HIVE's axis, and it scores as well as the robot in front. Two launch spots are also what lets partners launch at the same time.

At practice: find the nearest and the farthest distance that still scores, from straight on and from an angle, and mark the middle of each on your practice field. Teach your drivers those spots, not "somewhere near the goal".

3. Run an AUTO, and plan it with your partner

Two good solo routines collide at the same launch spot and score less than half of a planned pair. The value of AUTO is also more than its points: a planned pair starts TELEOP with full hoppers and a loaded CELL.

What made the pair work:

• Plan by start position. One routine for the robot that starts at the right, and one for the left. Each robot launches from its own spot, 0.62 m24 in. apart, so they never share one.
• Wait on a sensor, not on the clock. Each robot holds its fire until its camera reads the AprilTags of the raised CELL. An earlier version used clock waits, and a launch 1.5 s early cost 23 AUTO points.
• Put the short trip on the critical path. The slow trip fills the dead time while you wait for your partner. That order moved the third TIP from 21 s to about 17 s, which left time for a fourth.

Before a match: ask your partner where they start, where they launch from, and what they collect. Agree on who goes where.

4. Keep doing the thing that pays, and send at most one robot to the side objective

A TIP is 20 points for 6 to 8 elements. A finished FLOWER is 17 points for 6 elements, and it takes longer. One robot on FLOWERS for the last 30 s costs nothing and takes those FLOWERS away from the opponent. Two robots on FLOWERS for the last minute is the worst plan tested.

The answer changes if your launcher is poor. At 8 times the launch spread, FLOWER work from 45 s adds 33 points, because placing doesn't depend on aim.

In general: work out the points per element and per second for every way to score, and check it against how well your robot does each one.

5. Watch where your robot waits

Where a robot's 120 s of TELEOP went in these matches:

Driving is about 73% of the match, and launching is 5%. Fixing decisions, with no change to the robot, was worth about 32 points per alliance. The mistakes were ones that a drive team makes too:

• A long trip with a nearly empty hopper. A robot carried one element across the field and waited 5 s for a CELL.
• Driving past a free element. A robot with two empty slots passed a POLLEN 0.22 m8.7 in. away, because it had "enough".
• Assuming your partner has it. A robot left a CELL that needed one more element, because its partner held a full load, and the partner was still across the field. Call out who delivers first.
• Pushing your partner. Two robots arrived at each other's launch spots and shoved for 2 s. The closer robot takes the spot that it is on.

At practice: film a match, and count the seconds in which a robot holds elements and doesn't launch, waits for something, or drives past an element with room in the hopper.

6. Release your whole load fast, and face the shooter away from the intake

The one-at-a-time row is within the noise. Time at the launch spot is paid on every cycle, about 14 times per match. The HIVE also tips faster under a heavy overload: 1.2 s, against 1.7 s when the load barely crosses the threshold.

The shooter's side matters for a reason that carries to other games. With the shooter opposite the intake, the robot arrives at the launch spot from a pickup without turning around, and its intake faces the wall where the dumped elements land.

7. Build light, stay small, and don't chase top speed

Mass is the clear one: about 5 points per kilogramabout 2.4 points per pound, in a straight line. The field is 3.7 m12 ft across and crowded, so a robot spends its time speeding up and stopping, and traction limits how fast any robot speeds up. A light robot stops sooner, and a robot that settles sooner launches sooner.

Size pays only at the extreme, where a robot slips past its partner and through the HIVE frame. Gearing between 312 rpm and 700 rpm makes no measurable difference, and 850 rpm hurts.

The simulator doesn't check whether your mechanisms fit in 30 cm12 inches or weigh 7 kg15 lb. Treat the tables as the price of every pound and every inch.

8. Check what your intake can actually reach

Elements end up against walls and in corners. An intake that is narrower than the chassis can't reach an element on a wall while the robot drives along that wall. In this model the gap was 4.5 cm1.8 in. on each side, and a wall sweep missed the elements by 1 cm0.4 in. to 2 cm0.8 in..

An idea from one of our mentors fixed the wall problem without a new mechanism: turn the robot 30° toward the wall, put the front corner on the wall, and drive to the field corner. At that angle the wall crosses the mouth of the intake. One such sweep filled a hopper in 0.8 s, where a straight pass took 6 s to find two elements.

An intake on both ends lets the robot take an element with whichever end is nearer, so it turns less. It was worth 26 points in one test and about 10 in another. With intakes on both ends, the shooter's side stops mattering: 434 with the shooter at the front, and 438 at the rear. A rear intake has to share a face with a rear shooter, which is a packaging problem that the simulator ignores.

At practice: put elements against a wall and in a corner, and find out how your robot gets them.

9. Go where the elements end up, and let them land first

A TIP dumps 8 elements, and that dump is most of the supply for the next TIP. Three measurements changed how the robots collect:

• The dump lands in the same place every time: along the wall behind the launch spot. A robot that drove 25 cm10 inches into that wall after its last shot, and then launched again, gained about 12 points.
• The dump needs about 2 s to land. Robots that swept early found one or two elements. Robots that waited found three or four.
• Get out of the way. A robot that stands in the path of the dump blocks it. Backing under the HIVE let the dump pass overhead to the wall.

The supply is finite. Changes that collected more in AUTO raised AUTO points and barely moved the match score, because those elements weren't there in TELEOP. The changes that raised the match score removed waiting and driving.

In general: learn where game elements come to rest after every scoring event, and plan your routes to end there.

10. Know what your last 8 seconds are worth

The robots used to PARK whenever the clock said so. A robot holding four elements that would finish a TIP drove off to PARK for 5 points instead of 20.

Choosing by value was worth about 5 points per alliance, measured to within 1 point.

Ranking points change the answer. In a qualification match, the SWARM ranking point is worth more than 15 match points, so a robot whose PARK the alliance still needs for it must PARK. A normal AUTO already secures SWARM with LEAVE and AUTO PARK. In a playoff match, only points count, so take the TIP.

Before a match: decide who parks and who takes a last cycle, and what on the scoreboard changes that.

The build that these lessons point to

The tables change one thing at a time, so we tested whether the traits stack. Two robots with every design choice from lessons 6 to 8 played 24 matches against two baseline robots.

The combined build scored 508, and its baseline opponents scored 412. The margin is 96 points, with a standard error of 9. With baseline robots on both sides, the same seeds give 417 to 416. So the traits stack: the build is worth about 90 points per alliance, which is close to the sum of its parts.

How that alliance plays is lessons 2 to 5, 9, and 10: a planned AUTO pair, launches from the middle of the scoring window, both robots on the main cycle, routes that end where the dumps land, and a last launch over a PARK when the ranking points allow it.

Hold the build loosely, for four reasons:

• The launcher still comes first. The combined build keeps the baseline launcher. At 3 times the spread, lesson 1 takes away more than this build adds.
• Nobody plays defense here. A 7 kg15 lb, 30 cm12 in. robot gets pushed around in a way that these matches never test. That is the best reason for a real robot to carry more mass.
• Packaging is ignored. A four-element catapult, two intakes, and a rear shooter in 30 cm12 inches and 7 kg15 lb is a hard build. The tables price every pound and every inch. They don't say that the build is possible.
• The opponent is the baseline. Against the same build, the margin is whatever the better launcher and the better decisions give.

When to play defense

We tested one robot that defends for all of TELEOP. It goes after the opponent that carries the biggest load: it takes that robot's launch spot if it can get there first, and otherwise it drives into it. It breaks contact every 1.6 s, and no PIN was called in any match.

Defense rarely wins a match, for the reason that experienced teams give: the best defense against defense is a lead after AUTO. A robot with a poor TELEOP has a poor AUTO too. At 3 times the launch spread, AUTO points fall from 88 to about 43. A weak alliance that defends improves its margin and still loses every match, with a lower score for both sides. Between equal alliances, defense is a wash.

The exception is a mixed alliance, which is what a qualification schedule deals you. A strong robot with a weak partner lost all 24 matches while both scored, and won 15 of 24 with the weak partner on defense, from about 47 points behind after AUTO. Two things moved: the opponent lost 70 points, and the alliance's own score rose by 21. A poor launcher scatters elements and takes up the launch spot and the floor elements that its strong partner uses better.

Before a match: if your partner can't score reliably, ask them to defend, or at least to stay clear of your launch spot and your elements. If you are that partner, offer it.

A lighter version did nothing: a robot that shoves an opponent that is lined up to launch, and then launches itself, moved the margin by 3 points.

Two limits. The only contact rule in the simulator is the PIN rule, G421, so check your game's rules on contact near a scoring zone. And the opponents here don't adapt: a real drive team launches from another spot when a defender takes theirs.

What didn't help

• A shooter that launches from both ends. A two-position turret or a double catapult scored 412, where the baseline scored 417. A mecanum robot turns while it drives to the launch spot, or while it waits for the CELL, so the turn is already free. The dual intake helps because pickups come quickly and at any heading.

• Top speed. See lesson 7.
• Harder braking. Planning to stop at 5.0 m/s²16 ft/s², where the baseline plans 3.0, left the combined score of both alliances unchanged. At 4.0 m/s²13 ft/s², robots overshot the launch spot in one test.
• Wider or tighter lines around obstacles. A 4 cm1.6 in. and a 12 cm4.7 in. clearance, where the baseline is 8 cm3.1 in., both scored a little higher than the baseline, so there is no pattern. An earlier round of matches said that tight lines paid, and this round doesn't confirm it.
• Clever partner rules. "The nearer robot takes the better launch spot" lost 30 points per alliance, because the spot changed while a robot drove to it. Fixed roles with one narrow exception did better.
• Five AUTO TIPS. About half of each dump lands across the field center line, where G402 keeps you out in AUTO. Four was the most that any plan reached.
• A language model as drive coach. In six paired matches it scored 333, and a fixed policy scored 340. It wavered toward FLOWERS at 32% to 52% confidence, against a written strategy that said never.

What a simulator can't tell you

• Ball masses are assumed. FIRST doesn't publish them. The HIVE is calibrated to the six rows of the Event Field Setup Guide, section 12.3, so the counts that tip it are right. Bounce and roll are estimates, and where a dump lands depends on both.
• Nobody plays defense. No robot blocks, pins on purpose, or steals from an opponent's FLOWER.
• The drivers are code. The code never tires, and it still collides about 3 times per robot per match. A human driver gains more from a forgiving robot than this driver does: a wide intake, a wide scoring window, and a light chassis.
• Designs are parameters. A chassis of 30 cm11.8 in., a robot of 7 kg15 lb, and a catapult are numbers here. Whether you can build them is outside the model.
• The intake is a box. An element is collected when its center enters a box in front of the robot. The 30° wall sweep works because of that geometry. Test it on a real intake.
• The baseline moves. Twelve matches per row put a standard error of 5 to 12 points on every score in the tables.

How the matches were run

Every match is a full 2:30 MATCH with four robots in a rigid-body physics simulator built from the official field CAD. AUTO is a fixed script per robot that reads only its own pose, its hopper count, and the AprilTags of the HIVE. In TELEOP, a planner picks pickups, plans paths around the field elements and the other robots, and launches only when the predicted shot scores.

• 744 matches tested the designs and strategies in lessons 1 to 4 and 6 to 8: red gets the change, and blue is two baseline robots. The pooled tables are in experiments/run-v2-report.md.
• 708 matches tested one decision change at a time on 24 fixed seeds, for lessons 5, 9, and 10. The log is experiments/policy-loop.md.
• 24 matches tested the combined build. They are the row e24-meta-build-red-only in experiments/policy-loop.jsonl.
• 1,524 earlier matches used an older planner and a heavier robot. This round replaces them. Two of their results changed: gearing for speed no longer hurts up to 700 rpm, and the shooter's side now matters.

To rerun the design tables, run npx tsx scripts/exp-run.ts all 12 11 101 TAG with a tag of your own, and then run npx tsx scripts/exp-report.ts. To watch a match, run npm run dev. To configure a robot, click the gear icon next to its number.