Three HIVE TIPs. One autonomous robot.
One active autonomous robot reached three completed HIVE TIPs within 30 seconds while its alliance mate remained inactive. Across 20,000 recorded training episodes, the highest safety-qualified active-alliance count was three, appearing in three episodes. No four-TIP solo result appears in this snapshot. [1]
The retained Red replay for seed 32266098 completes its TIPs at 2.517, 15.200 and 28.492 seconds, scoring 70 points against the two scripted opponents' 63. This is a training-episode witness, not the measured reliability of a frozen solo policy or a physical robot. [2]
Watch one robot complete three AUTO TIPs
Original recorded seed 32266098: the active Red learner completes three TIPs and scores 70; the two scripted Blue opponents score 63 with two TIPs. Red robot 1 is labeled INACTIVE throughout. This is a retained training episode, not a newly simulated match.
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.
Two TIPs dominate the observed distribution
Active-alliance TIP counts in 19,999 geometry-valid episodes with no PIN or teammate-overrun flags. One flagged episode remains reported but is excluded here. Opponent-alliance outcomes are not included.
Make solo mean solo
The v12 experiment controls only one robot on the learning alliance. Its partner receives no drive, intake, shooter, expansion or passing actuation. The partner remains a passive collidable body, so it can be displaced by contact; it is not anchored or removed from the field. [2]
The active robot begins with four carried POLLEN. Exactly four existing teammate preloads rest on the floor touching the inactive robot's front edge. They are not free points or an automatic inventory transfer: the learner must physically collect them. The total modeled inventory stays at 56 pieces, and the 50% open-field AUTO pickup rejection is unchanged. [2]
The pinned TU03 rule review did not permit a starting robot in the LOADING ZONE. Both robots therefore use legal wall starts rather than an invented parked start. The inactive mate receives no initial PARK credit. Rule legality and real field placement must still be checked against the current manual. [2][4]
Two opposing scripted robots remain active. Episodes end at exactly 30 seconds with no TELEOP, ranking points or extra post-cutoff settling credit. Only completed TIP events strictly before 30 seconds enter the AUTO count. [2]
A rare frontier is not a dependable opening
Of the 19,999 safety-qualified training episodes, 34 recorded zero TIPs, 2,205 recorded one, 17,757 recorded two and three recorded three. The three-TIP seeds were 31263128 (Red), 32266098 (Red) and 32266139 (Blue). Only seed 32266098 retains a replay in the inspected snapshot; the other two counts are supported by persisted statistics, not reconstructed trajectories. [1][2]
The final generation-19 frozen evaluation averaged 55.167 AUTO points for the existing champion across 48 balanced-color outcomes. The three candidate mean gains were +0.229, -0.875 and -0.563 points; none passed the promotion gate. This evaluation is separate from the best exploratory training episode. [3]
Optimize the whole refill loop
Without an active partner, preload delivery, travel, gathering and the next volley compete for the same robot's time. That changes the experiment, so the two-robot four-TIP opening cannot simply be advertised as a solo capability. The evidence points to refill scheduling and reliable physical collection as questions to measure next, not to a proven solo ceiling. [1][2]
The October 5 study was paused after 20,000 episodes and its generation-19 evaluation. These posts report the preserved study; preparing this article does not resume training. Three is an observed possibility under these conditions. Repeatable physical three-TIP AUTO remains a separate validation task. [1][3]
The retained solo witness
Red seed 32266098. The third completion leaves about 1.508 seconds before the strict AUTO cutoff. The inactive teammate contributes no actuator commands.
Exact TIP completion times
| TIP | Completed at | Phase |
|---|---|---|
| 1 | 2.517 s | AUTO |
| 2 | 15.200 s | AUTO |
| 3 | 28.492 s | AUTO |
Training witnesses and retained evidence
Each count belongs to the single active learner's alliance. Missing recordings are explicitly marked; no playback or timestamps are invented.
| Seed | Alliance | TIPs | Alliance points | Event replay |
|---|---|---|---|---|
| 31263128 | Red | 3 | 69 | Not retained; statistics only |
| 32266098 | Red | 3 | 70 | Retained, hash checked |
| 32266139 | Blue | 3 | 68 | Not retained; statistics only |
Source [1].
What this study does not prove
- The frontier combines changing exploratory training actors; it is not a frozen-policy success-rate estimate.
- Two of the three frontier episodes lack retained recordings. Their counts are statistical evidence only.
- Idealized sensors, equal hardware, 2% conditional misses and 0.8-second HIVE movement remain uncalibrated.
- The inactive teammate's four floor preloads are conserved physical supplies, not automatic learner credit.
- No four-TIP solo result, proven physical maximum or hardware-ready autonomous routine 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.
- v12 solo all-episode snapshot
20,000 active-alliance statistics rows; 19,999 pass the stated safety filter. Includes the complete count distribution and all three frontier identities.
- Solo actor recording, seed 32266098
151 retained frames and exact completion events. Original compressed recording SHA-256: 0c46ff7c64d80d911cf48997f4ea35b342b5b2ca8dc03d3fb61349bbc28f2aae. Its pinned setup records the inactive teammate and conserved preloads.
- Generation-19 frozen promotion evaluation
Separate paired 48-outcome champion/candidate evaluation. Includes mean points, gains, rejection decisions and original evaluation-file hash.
- Official BIOBUZZ Competition Manual
TU03 was pinned on October 5 for G304.E, G304.G and section 10.3.4. The linked manual can be updated; consult its current revision before a physical test.