TransFuser on NAVSIM
Reproducing an end-to-end driving policy that fuses camera and LiDAR in bird's-eye view, trained on multiple GPUs on ASU's Sol cluster via SLURM.
The problem
End-to-end driving models map raw sensors straight to a plan, but published results are hard to reproduce. Without a verified baseline, no improvement can be trusted.
What I built
- Set up the NAVSIM benchmark pipeline and verified it against the official checkpoints and published scores
- Reproduced TransFuser's camera and LiDAR fusion in a shared bird's-eye-view representation
- Scaled training across multiple GPUs on ASU's Sol HPC cluster with SLURM job scheduling
Key decisions
Verify the pipeline with official checkpoints before training
It separates evaluation bugs from training bugs, so a mismatch has exactly one place to look.
Multi-GPU training on a shared cluster
Full-scale training is the only way to match published numbers rather than a scaled-down approximation.
Progress
0 of 4 milestones- Pipeline verified against official checkpoints
- Single-GPU training run
- Multi-GPU training on Sol
- Match published PDM score
Tech stack
PyTorchSensor fusionBEVNAVSIMSLURMMulti-GPU