MotionLM Processed WOMD Dataset

A processed, lightweight version of the Waymo Open Motion Dataset (WOMD), prepared for training a MotionLM model.

The original WOMD scenarios have been transformed into compact PyTorch tensor shards (.pt.zst, zstd-compressed) so they can be streamed directly into the training loop without the overhead of decoding protobuf scenarios on the fly.

Splits

Split	Shards	Size
training	1000	~114 GB
validation	150	~10 GB
test	150	~11 GB

File naming follows {split}.{i}of{n}.pt.zst.

Usage

Used to train the MotionLM implementation at https://github.com/Mol2017/MotionLM. See that repository for the data loader, model code, and training scripts.

import torch
import zstandard as zstd
import io

with open("training/train.0of1000.pt.zst", "rb") as f:
    decompressed = zstd.ZstdDecompressor().decompress(f.read())
batch = torch.load(io.BytesIO(decompressed))

Source

Derived from the Waymo Open Motion Dataset. Please consult the Waymo Open Dataset license for terms governing use of the underlying data.