File size: 2,724 Bytes
21c858b 157ff3c c171a3f 157ff3c |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 |
---
license: apache-2.0
pipeline_tag: reinforcement-learning
tags:
- rlgym
- rocket-league
- RLBot
- PPO
---
# CanoPy
CanoPy is a self-playing reinforcement learning Rocket League agent designed for the `RLBot Championship 2025`.
It uses PPO (Proximal Policy Optimization) to learn 2v2 gameplay through self-play. The agent is trained to play effectively on both blue and orange teams and can generalize to various team compositions.
## Model Details
- **Framework:** RLGym + RLBot v5
- **Algorithm:** PPO (via `rlgym-ppo`)
- **Team size:** 2v2
- **Action repeat:** 8
- **Observations:** `DefaultObs` with normalized positions, angles, velocities, and boost
- **Action space:** Lookup table actions with repeat frames
- **Reward shaping:** Combined reward including:
- Speed toward ball
- In-air bonus
- Ball velocity toward goal
- Goal scoring reward
## Training Configuration (from `config.json`)
- **Number of processes:** 4
- **Minimum inference ratio:** 80%
- **Steps per checkpoint:** 1,000,000
- **PPO batch size:** 100,000
- **PPO minibatch size:** 50,000
- **PPO epochs per update:** 2
- **Experience buffer size:** 300,000
- **Policy network layers:** [256, 128]
- **Critic network layers:** [256, 128]
- **Policy learning rate:** 0.0001
- **Critic learning rate:** 0.0001
- **PPO entropy coefficient:** 0.01
- **Standardize returns:** true
- **Standardize observations:** false
- **Total training steps:** 1,000,000,000
- **Checkpoint directory:** ./checkpoints
## Intended Use
CanoPy is intended for research, competition, and experimentation within the RLBot framework. It is designed to compete in the ML bot bracket of the RLBot Championship 2025.
## Limitations
- Performance is dependent on training; untrained or partially trained models may perform poorly.
- The bot has been trained for standard Rocket League 2v2 matches; it may not generalize to unusual map sizes, mutators, or game modes.
- Does not include human-like strategy beyond what PPO has learned from self-play.
## Evaluation
CanoPy can be evaluated using the `evaluate()` function in the training script. Expected evaluation includes average episode returns and gameplay against copies of itself.
- **Note:** To meet RLBot Championship submission requirements, further testing against Psyonix Pro bots may be necessary.
## Contact / Author
- **Author:** FlameF0X /// Discord handler `@flame_f0x`
- **Competition:** RLBot Championship 2025
|