ParticleSplat checkpoints
Checkpoints for ParticleSplat: Self-supervised Object-centric Latent Particle Splatting (project page, code). ParticleSplat decomposes a scene into a small set of latent particles from a few posed RGB-D views and renders them with feedforward 3D Gaussian splatting; the particles also serve as tokens for imitation learning.
Representation models
One model per benchmark. Every directory is a complete training run: config.yaml (merged configuration),
config.json, train.log and checkpoints/model.pth.
| Directory | Setting | Training config in the code repository |
|---|---|---|
rlbench_multiview_particlesplat/ |
RLBench, 2 × RGB-D context views, all ten tasks | configs/experiments/rlbench_multiview_particlesplat.yaml |
rlbench_singleview_particlesplat/ |
RLBench, 1 × RGB-D context view (front camera), all ten tasks | configs/experiments/rlbench_singleview_particlesplat.yaml |
mimicgen_particlesplat/ |
MimicGen, 2 × RGB-D policy cameras, all twelve tasks | configs/experiments/mimicgen_particlesplat.yaml |
huggingface-cli download lyuxinghe/particlesplat --include "rlbench_multiview_particlesplat/*" --local-dir checkpoints
Policies
One EC-Diffuser policy per task (200k steps) on frozen ParticleSplat tokens, under policy/rlbench/<task>/ and
policy/mimicgen/<task>/. Each directory holds ckpt/state_399_step200000.pt, args.json, the EC-Diffuser
*_config.pkl files and diff.txt. Training and rollout evaluation are documented on the
policy-learning branch of the code repository.
huggingface-cli download lyuxinghe/particlesplat --include "policy/rlbench/close_jar/*" --local-dir checkpoints
Data
lyuxinghe/particlesplat-rlbench: multi-view RLBench episodes and policy artifacts.lyuxinghe/particlesplat-mimicgen: multi-view MimicGen renders, demonstrations and policy artifacts.
The code is released under the MIT license.