← sources

AI Beats the Human World Record on One of the Hardest Surf Maps

source: reddit (reddit:reinforcementlearning)published 2026-10-24 15:39ingested 2026-10-25 02:39

retained text

A reinforcement learning agent completes one of the hardest surf maps (surf\_src\_sidistic) faster than the human world record.  
  
The world-record demo by player krs was taken from [Surf Gateway](https://speedrun.eu/maps/surf_src_sidistic/records)  
  
The agent’s camera movements are smoothed to make the video more pleasant to watch. Because of this, you can’t see its high-frequency micro-strafing, but you can follow the A/D inputs. Without smoothing, the movement looks very shaky. These are movements no human could execute.  
  
The agent was trained without any human data. This is pure reinforcement learning, with no imitation learning.

The agent is trained in optimized simulation that inherits CS 1.6 movement and surfing logic. The algorithm is PPO with a lot of small additions. After the agent finished the map, I additionally improved the policy by running AlphaZero style expert loop (search + distillation into policy). 

  
Repository: [https://github.com/Sorrow321/RL\_CS1.6\_Surf](https://github.com/Sorrow321/RL_CS1.6_Surf)  
  
Music: Utah Saints - Something Good '08 (High Contrast Remix)

Video on [YouTube](https://www.youtube.com/watch?v=vZkLsSVVvAM)

document fetched 2026-10-25 00:34 · canonical https://reddit.com/r/reinforcementlearning/comments/1xebtu4/ai_beats_the_human_world_record_on_one_of_the