Qwen3.5-9B-aerialsim-rl-step100

Qwen3.5-9B fine-tuned with GRPO on the aerialsim-env 17-family aerial safety scenario mix, trained via prime-rl on a 4x H100 pod. This is the step-100 checkpoint from run smoke_17family_qwen35_9b_4gpu_wide_2026-07-15.

Training summary

• Base: Qwen/Qwen3.5-9B (dense multimodal, qwen3_5 arch)
• Environment: aerialsim-env v0.1, Cosys-AirSim / UE5.5 Blocks scene, 17 scenario families
• Observation mode: both (RGB frame + text scaffold with landing_zone_clear, obstacle_ahead, distance_to_goal, waypoints)
• Reward: binary success on terminal action
• Algorithm: GRPO, batch_size=4, group_size=4
• Optimizer: AdamW with CPU offload, lr=5e-7
• Rollout sampler: temperature=0.3, max_completion_tokens=1024
• Eval sampler: temperature=0.0
• Sequence length: 4096
• GPU split: 2 train + 2 infer, tensor_parallel=2 on each side
• Steps: 100

Evaluation

Per-family deterministic (T=0, single decode) eval with scaffolded text state is pending; will be reported in the accompanying paper.

Intended use

Research on visual-grounded aerial safety decisions and RL fine-tuning of multimodal LLMs on simulated aerial mission scenarios.

Author

Nahid Alam (@nahidalam) — Maya Multimodal