lyuxinghe/particlesplat-mimicgen

ParticleSplat: MimicGen data

Multi-view MimicGen data used by ParticleSplat: Self-supervised Object-centric Latent Particle Splatting (project page, code).

Twelve tasks (stack, stack_three, square, threading, coffee, three_piece_assembly, hammer_cleanup, mug_cleanup, kitchen, nut_assembly, pick_place, coffee_preparation), 200 D0 demonstrations each. Every timestep of the renders provides 20 posed RGB-D views at 128×128: slots 0–7 form a camera path from the agentview to the sideview camera (the two policy cameras), slots 8–19 a look-at arc used as reconstruction target. Pose files are OpenCV camera-to-world matrices and depth is 24-bit packed over [0.01, 10] m.

Layout

Path	Content
train_data/<task>_d0.tar, test_data/<task>_d0.tar	one archive per task and split; extracts to <split>/<task>_d0/{all_variations/episodes/episode<N>/, rig_manifest.json, demo_index.json, depth_stats.json}
demos/<task>_d0.hdf5	the D0 demonstrations as robomimic hdf5 files with recorded simulator states
policy/<task>/tokens/<task>_ps_dlpgs_2view_tokens.pkl (+ .meta.json)	ParticleSplat tokens of the demonstrations for policy learning
policy/<task>/latents/<task>_ps_dlpgs_latents.npz	encoder latents
policy/<task>/frames/{cameras.json,manifest.json}	policy camera poses and frame provenance

huggingface-cli download --repo-type dataset lyuxinghe/particlesplat-mimicgen --include "*_data/*.tar" --local-dir data/mimicgen
for f in data/mimicgen/*_data/*.tar; do tar -xf "$f" -C data/mimicgen && rm "$f"; done

Representation training and evaluation are documented in the code repository; policy learning on the policy-learning branch. The demonstrations come from MimicGen and are distributed under its terms.