rcd-ogbench-checkpoints

RCD / CompDiffuser Pretrained Diffusion Checkpoints (OGBench)

Pretrained compositional diffusion planner checkpoints used in the RCD experiments, trained with the CompDiffuser codebase on OGBench environments: PointMaze, AntMaze, HumanoidMaze, AntSoccer, and Cube.

Only the final training checkpoint of each run is included (plus the small config/metadata files saved next to it), so the download stays manageable.

Download

pip install -U huggingface_hub

# Everything (~17 GB)
hf download leekwoon/rcd-ogbench-checkpoints --local-dir ./rcd_checkpoints

# Or a single environment, e.g. HumanoidMaze Medium only
hf download leekwoon/rcd-ogbench-checkpoints \
    --include "rcd_ogbench/logs/humanoidmaze-medium-stitch-v0/*" \
    --local-dir ./rcd_checkpoints

# Integrity check (optional)
cd rcd_checkpoints && md5sum -c checksums.md5

(Older CLI: huggingface-cli download leekwoon/rcd-ogbench-checkpoints --local-dir ./rcd_checkpoints)

Contents

Each planner / inverse-dynamics run directory also contains args.json, model_config.txt, and the *.pkl config files saved by the trainer.

PointMaze uses a simple waypoint controller instead of a learned inverse-dynamics model, so only the planner checkpoint is provided for it (no *_invdyn_* run). AntMaze, HumanoidMaze, and AntSoccer each ship a planner plus a learned inverse-dynamics model. These planners are our own runs trained with the CompDiffuser-style codebase; loading a different (e.g. upstream CompDiffuser) checkpoint into these configs will not reproduce the RCD numbers, because the normalizer statistics and diffusion/overlap hyper-parameters are tied to the specific training run.

Usage

The directory layout mirrors the logs/ layout of the CompDiffuser-style codebase, so you can merge it directly into your repository root:

# Locomotion (HumanoidMaze, AntSoccer)
cp -r rcd_checkpoints/rcd_ogbench/logs/* <your_locomotion_repo>/logs/

# Manipulation (Cube)
cp -r rcd_checkpoints/rcd_ogbench_manipulation/logs/* <your_manipulation_repo>/logs/

The eval configs load checkpoints with diffusion_epoch: 'latest', which resolves to the included state_*.pt automatically. For example:

# HumanoidMaze Medium rollout
python diffuser/ogb_task/ogb_maze_v1/plan_ogb_stgl_sml.py \
    --config config/ogb_hum_maze/og_humM_Me_o2d_Cd_Stgl_PadBuf_Ft64_ts1k_h336_ovlp128_ovdm5_ts512_bs192_bd128_td96_drop02.py

# AntSoccer Arena rollout
python diffuser/ogb_task/ogb_maze_v1/plan_ogb_stgl_sml.py \
    --config config/ogb_ant_soc/og_antSoc_Ar_o17d_DiTd768_PadBuf_Ft64_ts512_fs4_h160_ovlp56MditD384.py

Cube low-level (DQL) policy

Cube rollouts use the public MCTD (cube branch) Diffusion-QL performer as the low-level controller. The epoch-200 checkpoints (the default --dql_epoch 200) are included under rcd_ogbench_manipulation/dql_results/<env>/.

The eval script (eval_cube_multi_mctd_rollout.py) expects them under external/mctd_cube_public/dql/results/<run_name>/, where the original run name is

<env>|exp|diffusion-ql|T-5|lr_decay|ms-offline|k-1|0|3|1.0|False|cql_antmaze|0.2|4.0|10

(we renamed the folders here because | is not portable). Either restore that folder name, or simply pass --dql_dir <path> to the eval script.

Citation

If you use these checkpoints, please cite the RCD paper and CompDiffuser (arXiv:2503.05153).