How many HIVE TIPs are possible?
Our HIVE-focused Reinforcement Learning Environment (RLE) study reached 13 completed TIPs for one alliance in a full simulated match. That is the highest safety-qualified count in this study's 50,000 recorded matches, not a mathematical ceiling or a claim about a physical Cookie Coders robot. [1]
The important result is not just the headline. All seven 13-TIP outcomes came from the scripted opponents, while the learning alliance reached at most 12 in safety-qualified training matches. The best witnessed sequence is evidence that the model can reach 13; it does not establish a learned policy that reliably does so. [1]
Watch the 13-TIP match
Original recorded seed 32261043: scripted Red finishes with 13 TIPs and 271 points; the Blue training alliance finishes with nine TIPs and 186 points. Playback includes the recording's passive settling after 150 powered seconds. It does not rerun the model.
Loading the recorded match...
[*] Starting score: Each alliance begins with 10 modeled field-content points: 4 GARDEN points (four starting POLLEN at 1 point each) plus 6 HIVE-retained points (three starting NECTAR at 2 points each). These are already present at setup, not earned by robot actions. We keep the original recorded totals; contents points change as pieces move and are not a fixed 10-point bonus. Replay evidence.
Starts paused at 1x. Positions are original recorded frames, not interpolated or resimulated. Exact TIP event timestamps can fall between frames; event jumps show the first recorded frame at or after completion.
Playback provenance
Original recording information appears when playback loads.
Download the sanitized replay dataSource [2]. Uncalibrated simulation; no claim of physical robot performance.
The frontier is not the typical result.
TIP counts for both alliances in 46,594 safety-qualified matches: 93,188 alliance outcomes. This is a changing training population, not a frozen-policy reliability benchmark.
A controlled question, not unlimited scoring
The v10 experiment focuses on HIVE collection, extraction and shooting. FLOWER scoring, GARDEN delivery, free roaming and defensive actions are excluded from its offensive action mask. Robots still need to gather conserved pieces, navigate, aim and physically fill the HIVE; commands do not teleport balls into the score. [2]
The model runs 30 seconds of AUTO followed by 120 seconds of TELEOP actuation. These 150 powered seconds are not an official wall-clock match schedule: the official eight-second transition is not represented here. Recorded v10 outcomes can include passive settling after powered play; the illustrated thirteenth completion is at 150.000 seconds, not a post-cutoff bonus. [2][4]
For this article, safety-qualified means the whole match is geometry-valid with zero modeled PIN flags and zero teammate-overrun flags. We exclude 3,406 matches from the frontier chart but retain their existence in the denominator. This filter is not a full referee audit or a physical safety certification. [1]
Starting position changes the available work
In a separate 12-seed scripted comparison, a mirrored opposite-wall start raised average TIPs per alliance from 9.542 to 10.375. Sixteen of the 24 paired alliance outcomes improved, seven tied and one regressed. Modeled PIN flags fell from two to zero. The comparison includes all outcomes in each arm rather than removing the baseline's flagged cases. [3]
The split gives one preloaded scorer access to the first HIVE face and positions its partner across the field for the next volley. The engineering lesson is to overlap useful work: one robot reloads while its partner addresses the next scoring opportunity. The paired probe reached at most 12 TIPs; the later 13-TIP witnesses belong to the larger training study, not that probe. [2][3]
Why 0.8 seconds is not a repeat-TIP cycle
The estimated 0.8-second damper parameter measures HIVE movement after a TIP begins. It excludes collection, driving, aiming, shooting, ball flight and refilling. Dividing match time by 0.8 therefore does not produce a defensible TIP maximum. [2]
A true physical upper bound has not been derived. Thirteen is an observed frontier under the stated assumptions, and any real-world claim needs timed HIVE motion, measured hardware, representative opponents and repeatable physical trials. Higher scoring and parking objectives also need separate evaluation; this post does not claim 13 TIPs and six ranking points together. [1][2]
Inside a 13-TIP match
Red, seed 32261043: three AUTO completions and ten TELEOP completions. The thirteenth finishes at 150.000 seconds. Each marker is a completed event, not a requested shot.
Exact TIP completion times
| TIP | Completed at | Phase |
|---|---|---|
| 1 | 2.542 s | AUTO |
| 2 | 13.158 s | AUTO |
| 3 | 25.050 s | AUTO |
| 4 | 34.692 s | TELEOP |
| 5 | 44.958 s | TELEOP |
| 6 | 55.192 s | TELEOP |
| 7 | 67.758 s | TELEOP |
| 8 | 84.592 s | TELEOP |
| 9 | 99.183 s | TELEOP |
| 10 | 114.200 s | TELEOP |
| 11 | 123.833 s | TELEOP |
| 12 | 137.567 s | TELEOP |
| 13 | 150.000 s | TELEOP |
Two replay-backed frontier witnesses
Both scoring alliances use scripted controllers. Points belong to the alliance; they are not per-robot points or combined Red-plus-Blue totals.
| Seed / alliance | AUTO | TELEOP | Total TIPs | Alliance points |
|---|---|---|---|---|
| 32261043 / Red | 3 | 10 | 13 | 271 |
| 31263062 / Blue | 2 | 11 | 13 | 266 |
Source [2].
What this study does not prove
- The 2% intrinsic miss probability applies only to otherwise clear shots; blocked or badly aimed shots remain failures.
- Sensors, equal-capability hardware and the 0.8-second HIVE movement are idealized or uncalibrated.
- The 50% open-field AUTO POLLEN pickup rejection is a model assumption, not a measured hardware failure rate.
- Seven frontier outcomes among 93,188 safe alliance outcomes do not establish a fixed policy's success probability.
- No mathematical maximum, calibrated physical result or 13-TIP learned champion is established.
Where to play BIOBUZZ
BloxBuzz is where you can play BIOBUZZ; the RLE is where our agents train and are evaluated. The RLE does not run in Roblox, and our goal is for it to use the same game engine, rules and physics as BloxBuzz. These results come from the RLE, not from BloxBuzz gameplay, and this study does not verify that the two match.
Opens Roblox in a new tab. Roblox's age requirements and privacy practices apply. BloxBuzz is unofficial and is not affiliated with or endorsed by FIRST or RTX.
Play BloxBuzz on RobloxSources and evidence
Linked JSON files are sanitized research extracts with sample counts, event witnesses and source hashes. Embedded 2D playback uses exact retained frame coordinates and TIP events; the full original actor logs remain privately retained. These extracts are not a complete reproduction package.
- v10 all-match snapshot
50,000 statistics rows; whole-match safety filter; both-alliance histogram and controller attribution. Snapshot: October 8, 2026.
- Retained v10 event witnesses and assumptions
Seven hash-checked actor recordings. Seeds 32261043 and 31263062 illustrate the 13-TIP frontier. The extract includes exact completion times and original recording SHA-256 values.
- Matched starting-layout control
Twelve identical seeds, 24 alliance outcomes per arm; standard versus opposite-wall start. Includes per-seed counts, validity, PIN flags and source-file hashes.
- Official BIOBUZZ Competition Manual
Use the current FIRST manual for timing, starting constraints and event scoring. Simulator compliance is not an inspection approval.