StageVLN-v0-r2r-uniform4-4gpu

R2R action-prediction fine-tune of Qwen/Qwen3.5-4B. The training set has 631,244 action states from 10,819 episodes. Each example uses the current RGB frame and up to four uniformly spaced preceding RGB frames. The model predicts one of MOVE_FORWARD, TURN_LEFT, TURN_RIGHT, or STOP.

Training: one epoch, global batch 64, bf16, FlashAttention 2, accelerated gated linear attention, and ZeRO-1 on 4 H100 GPUs. Base Git revision: fd072e9f3438f16099f46655009a3b91dcb6e971. The R2R validation/test sets and simulator were not used to train this export.