Final Results

Rankings of both tracks on the private test set, published October 17, 2026, updated November 4, 2026.

Every team listed here presents a poster at the competition's session at NeurIPS 2026 in Sydney on December 11.

Track 1 · Sim2Real

PlaceTeamTotal scoreStd. dev.
1np-user82.95550.0281
2Pistachio (roysegal)82.90060.0166
3PhysicsOracle82.60850.0045
4syouya_tobita82.56080.0223
5Aminos (doomduke2)82.48770.0173
6=skabob82.44340.0747
6=AxisTilted2 (spsanps)82.43980.0319
8iapetos191882.41410.0226
9JLShen (junlong)82.30550.0061
10IKNO (g2404426g)†82.12520.0512
11Math4AI (simonkempf)*81.84060.1064

Track 2 · LTTTA

PlaceTeamTotal scoreStd. dev.
1nmucke83.05290.0035
2Pistachio (roysegal)83.02150.0099
3=AxisTilted2 (spsanps)82.96010.0174
3=np-user82.95600.0226
5seantang82.83150.0089
6Separation Anxiety (oroikono)82.80310.0193
7Aminos (doomduke2)82.78630.0151
8xtgg23382.72100.0041
9=AI4Silence (gaswang)82.62710.0280
9=Cupid (andychang)82.61040.0653

= marks a shared place. † and * are explained under Teams below.

How the scores were obtained

Teams
Among the teams that submitted a valid Decision Phase training package (training code and Docker image) and remained eligible under the competition rules, each track ranks the 10 placed highest on that track's Development Phase leaderboard. Teams are listed by the team name they gave with their training package, followed by their Codabench account in parentheses where the two differ.

† The organizers missed the on-time Track 1 training package of IKNO (g2404426g). That package was trained and evaluated in November 2026 with the same procedure and added to the ranking, so Track 1 lists 11 teams.

* Math4AI (simonkempf) was evaluated in October as the next team on the reserve list, in the place left by IKNO's missed package. Its result stands.

Training
Each team submitted a separate training package for each track it competes in. The organizers trained each package from scratch three times on the organizers' NVIDIA A800 GPUs, which gives three trained models per team and track.
Evaluation
Each trained model was evaluated five times on the organizers' NVIDIA A800 GPUs, the same GPU type as for the Development Phase leaderboard. The evaluation followed the Development Phase protocol, run on the private test set instead of the Development Phase evaluation data. The five subscores were combined into a total score in the same way as for the leaderboard's final score.
Total score
The table shows the mean total score over the team's 15 evaluations in that track (3 trained models × 5 evaluations each) and the standard deviation over the same 15 evaluations. Higher is better.
Shared places
When two adjacent teams' mean scores differ by less than both teams' standard deviations, the evaluation cannot separate them, and they share the place, marked with =. Three places are shared:
  • Track 1, 6th place: skabob and AxisTilted2 (spsanps).
  • Track 2, 3rd place: AxisTilted2 (spsanps) and np-user.
  • Track 2, 9th place: AI4Silence (gaswang) and Cupid (andychang).
If a shared place carries a cash prize, the teams that share it split the prize equally.

What each place receives is listed under Prizes. Questions about the results: realpde-competition@googlegroups.com.