Pre-trained Checkpoints

Download the LiDAR-only checkpoints from HuggingFace Hub:

# Create checkpoints directory
mkdir -p checkpoints

# Download encoder checkpoint (9.6 MB)
huggingface-cli download sangramrout/topodiffuser encoder_lidar_only_best.pth --local-dir checkpoints

# Download diffusion checkpoint (4.0 MB)
huggingface-cli download sangramrout/topodiffuser diffusion_lidar_only_best.pth --local-dir checkpoints

Or use Python:

from huggingface_hub import hf_hub_download

encoder_path = hf_hub_download(
    repo_id="sangramrout/topodiffuser",
    filename="encoder_lidar_only_best.pth",
    local_dir="checkpoints"
)

diffusion_path = hf_hub_download(
    repo_id="sangramrout/topodiffuser",
    filename="diffusion_lidar_only_best.pth",
    local_dir="checkpoints"
)

Model Details:

• Encoder: LiDAR-only (3-channel: height, intensity, density) trained on sequences 00, 02, 05, 07
• Diffusion: Diffusion model with frozen encoder for trajectory prediction
• Training Date: February 17, 2026

Diffusion Training

python -u train_diffusion_only.py \
    --encoder_ckpt checkpoints/encoder_lidar_only_best.pth \
    --resume checkpoints/diffusion_lidar_only_best.pth \
    --epochs 500 \
    --batch_size 64 \
    --lr 1e-4 \
    --noise_schedule cosine

Encoder training

• Paper: out of 00-10 seq, only a handful was selected
• Current implementation: Entire dataset of seq 00-10- randomly selected training and validation set from the dataset.

Training on paper vs implementation metrics

all sequences

training only on lidar data

From Table I in the paper:

Metric	Paper Target	Your Best	Gap	Progress
minADE ↓	0.26 m	0.883 m	+0.62 m	70% worse
minFDE ↓	0.56 m	-	-	-
HitRate ↑	93%	82.5%	-10.5%	89% there
HD ↓	1.33 m	1.562 m	+0.23 m	85% there

Notes:

• minADE: Minimum Average Displacement Error
• minFDE: Minimum Final Displacement Error
• HitRate: Percentage of predictions within threshold
• HD: Hausdorff Distance

All three modalities for each sequence side by side