Fight win rate
57.1%848 wins out of 1,484 fights
A learning agent that uses recorded fights to improve movement, attacks and camera control. Explore the published results and how the project works.
Recorded fights against real players on public servers, across model versions. This is not a current-model benchmark.
Showing the latest saved sync · 2026-11-09
848 wins out of 1,484 fights
848 wins / 636 losses
Valid, confirmed combat outcomes
Rejected records and walkovers are excluded. Only aggregate results are public; player identities, recordings and model files remain private.
The five latest candidates, oldest to newest. Results belong to the exact model that played each fight.
| Model | Status | Detected? | Focus / goal | Real-player fights | Record | Win rate | W/L | Local arena |
|---|---|---|---|---|---|---|---|---|
| Beta 6 · Candidate 4 | Retired | ✕ | Improve prediction | 227 | 152 W / 75 L | 67.0% | 2.03 | No arena results |
| Beta 6 · Candidate 5 | Retired | ✓ | Improve movement | 432 | 301 W / 131 L | 69.7% | 2.30 | No arena results |
| Beta 6 · Candidate 6 | Retired | ✕ | Camera and accuracy | 189 | 93 W / 96 L | 49.2% | 0.97 | No arena results |
| Beta 6 · Candidate 7 | Retired | ✕ | Reward-driven combat without script assists | 141 | 64 W / 77 L | 45.4% | 0.83 | No arena results |
| Beta 6 · Candidate 8 | Current | ✕ | Not specified | 0 | No real-player fight results | — | — | No arena results |
Focus / goal describes the intended objective, not a guaranteed outcome. Detection refers to Minemen Club: ✓ detected, ✕ not detected. Current is the model selected for use. Retired models were promoted or used in real-player fights. Abandoned models never reached that stage; local-arena tests alone do not count. Statuses reflect the latest sync and remain visible when updates pause. Only qualifying recorded fights count. Local-arena results are shown separately and do not inflate real-player fight totals. Models with no results in a category have no measured win rate.
Watch EVO in action, with the version behind each clip.
EVO is an independent learning agent for Minecraft combat. It observes the fight and uses a trained policy to choose movement, attacks and camera actions.
The current pipeline combines demonstrations, opponent observations and reward-weighted offline learning. Updates are trained offline from recorded experience. Reinforcement learning from the agent’s own actions remains a research goal.
The agent reads movement, spacing and combat context. A learned policy chooses its next action, while the recorder captures what happened.
Recorded fights are checked for quality. Outcomes and performance signals weight the experience used to train candidates offline.
Candidates are compared with their parent model and tested in controlled play. A new model reaches production only after review.