Releases·Tool Release
ClashRoyaleAi: an open-source, deterministic Clash Royale simulator for RL, with recurrent PPO, lookahead search and expert iteration [P]
A developer released ClashRoyaleAi, an open-source headless C++17 recreation of Clash Royale's battle engine (with pybind11 bindings) serving as a Gymnasium RL environment, since the game has no API. The project includes a recurrent PPO agent trained with a simulation-based teacher curriculum and a self-play league. Early results show large gains from lookahead search, though the agent remains weak overall.
2 engagement1 sourceSun, Sep 27, 2026, 12:30 PM
AI Overview
- Headless C++17 simulator (pybind11, Gymnasium API) covers all 132 cards plus 41 Evolutions as of July 2026, runs ~10 ms per match
- Observation is a 13,977-float vector including opponent card history; actions pick a card plus one of 612 cells with legality masks
- Recurrent PPO agent (CNN+LSTM, 1.85M params, CPU-only) trains via curriculum against a simulation-ranking teacher piloting 16 ladder meta decks plus PFSP self-play
- 1-ply lookahead lifts win rate from 0.625 to 0.944 (160 paired matches, 95% CI +0.24 to +0.40); distilling search back adds +0.045 (1,600 matches, p=0.007)
- docs/DECISIONS.md logs reversals including a reward hack, a biased self-play null, and a discount horizon shorter than the match; from-scratch run in progress
Related
Products·launched
DelveRL: Open-source roguelike for training game-playing agents
DelveRL is an open-source, human-playable roguelike game designed specifically for training game-playing agents. It provides a structured API, deterministic simulation, procedural levels, and partial observability, enabling agents to explore, manage resources, fight enemies, and escape floors. It solves the problem of integrating games with agent harnesses, offering a local, self-contained environment for reinforcement learning research.
DelveRL is an open-source, human-playable roguelike game designed specifically for training game-playing agents. It provides a structured API, deterministic simulation, procedural levels, and partial observability, enabling agents to explore, manage resources, fight enemies, and escape floors. It solves the problem of integrating games with agent harnesses, offering a local, self-contained environment for reinforcement learning research.
17 engagement1 sourceredditSat, Aug 22, 2026, 05:32 PM
Releases·Tool Release
BVR Sim: An Open and High-Throughput Environment for Heterogeneous Air-Combat Reinforcement Learning
Researchers released BVR Sim, an open-source Gymnasium-style environment for beyond-visual-range air-combat reinforcement learning. It supports multiple JSBSim aircraft models with configurable weapons, sensors, and controllers, aiming to accelerate research in this challenging domain.
Researchers released BVR Sim, an open-source Gymnasium-style environment for beyond-visual-range air-combat reinforcement learning. It supports multiple JSBSim aircraft models with configurable weapons, sensors, and controllers, aiming to accelerate research in this challenging domain.
0 engagement1 sourcearxivWed, Aug 26, 2026, 06:21 AM
Products·building
Pokelike.xyz transformed into LLM and RL benchmark
A data scientist built an environment around Pokelike.xyz, a Pokémon-like game, to train reinforcement learning agents and benchmark LLMs. It provides a documented repo for others to create bots and test AI capabilities.
A data scientist built an environment around Pokelike.xyz, a Pokémon-like game, to train reinforcement learning agents and benchmark LLMs. It provides a documented repo for others to create bots and test AI capabilities.
2 engagement1 sourceredditWed, Aug 19, 2026, 11:10 AM
Products·concept
Teaching Neural Nets to Fight with RL
A project training two reinforcement learning agents to play a Street Fighter-like game to observe emergent fighting behaviors. It documents reward hacking problems and shows how reward shaping and league play produce interesting play. It serves readers interested in practical RL experiment design and write-ups.
A project training two reinforcement learning agents to play a Street Fighter-like game to observe emergent fighting behaviors. It documents reward hacking problems and shows how reward shaping and league play produce interesting play. It serves readers interested in practical RL experiment design and write-ups.
1 engagement1 sourceredditSun, Sep 27, 2026, 03:10 AM
Developments·Paper
CAST: Game Solvers as Turn-Level Teachers for LLM Agents
A new arXiv paper proposes CAST, a method that uses game solver state-value changes to provide turn-level credit assignment for training LLM agents in long-horizon games. This addresses the sparse-reward problem in RLVR by offering dense process signals without the cost of existing alternatives.
A new arXiv paper proposes CAST, a method that uses game solver state-value changes to provide turn-level credit assignment for training LLM agents in long-horizon games. This addresses the sparse-reward problem in RLVR by offering dense process signals without the cost of existing alternatives.
0 engagement1 sourcearxivTue, Jul 28, 2026, 05:39 AM