FreeFire aim agent (Phase 1) — amer224/freefire-aim-ai

Pixels (84x84) → aim deltas + fire decision. NatureCNN actor-critic (~1.7M params) trained with BC warm-start + PPO + curriculum in a custom Free-Fire-like aim simulator. Beats nothing yet in real games — this is research scaffolding, and the numbers below are from the controlled simulator on unseen seeds, not live play.

Latest training tail

• upd5020 g5140480 L4 HS/hits=nan% acc=nan% kill=nan%
• upd5021 g5141504 L4 HS/hits=nan% acc=nan% kill=nan%
• upd5022 g5142528 L4 HS/hits=nan% acc=nan% kill=nan%
• upd5023 g5143552 L4 HS/hits=nan% acc=nan% kill=nan%
• upd5024 g5144576 L4 HS/hits=nan% acc=nan% kill=nan%

Files

• best.pt — best-by-unseen-HS/hits checkpoint (model state_dict + meta)
• latest.pt — most recent checkpoint (resume: model, optimizer, scaler, global_step, level)
• config.yaml, metrics.csv — exact config + full log

Resume / eval

pip install -r requirements.txt
PYTHONPATH=. python3 -m src.train --config configs/base.yaml   # auto-resumes
PYTHONPATH=. python3 -m src.evaluate --ckpt best.pt --level 1 --episodes 100