Qwen3.5-4B-aerialsim-rl-step100
Qwen3.5-4B fine-tuned with GRPO on the aerialsim-env 17-family aerial safety scenario mix, trained via prime-rl on a 2x H100 pod. Step-100 checkpoint from run smoke_17family_qwen3_5_4b_bugfix_lowlr_2026-07-12 on the fixed env (Option A + prior segment_index / eval-prompt fixes).
Training
- Base: Qwen/Qwen3.5-4B
- Env: aerialsim-env with 17 scenario families, Cosys-AirSim / UE5.5 Blocks scene,
obs_mode="both"(image + text state), binary reward - GRPO, batch=8, group=4, lr=5e-7, T=0.3 train / 0.0 eval, seq_len=8192
- 100 training steps
Per-family eval (T=0, n=8 per family, deterministic single decode)
15/17 families perfect. patrol_with_reroute also 8/8 (RL closes the last 2 fails from cold-4B's 6/8). Only multi_obstacle_dense remains at 3/8 due to oracle noise (heuristic policy scores 5/8 on the same family).
Aggregate: 131/136 = 0.963 (matches the heuristic policy ceiling).
Author
Nahid Alam (@nahidalam) - Maya Multimodal.