VAM robomimic assets — everything the robomimic runs need, in the dev/dppo_assets layout
Download into the repo's asset root and every config path resolves:
huggingface-cli download knightnemo/vam-robomimic-assets --local-dir dev/dppo_assets
export DPPO_LOG_ROOT=$PWD/dev/dppo_assets
export DPPO_DATA_DIR=$DPPO_LOG_ROOT
export DPPO_EXPERT_ROOT=$PWD/dev/dppo_assets/experts
| file | what | md5 | bytes |
|---|---|---|---|
robomimic-pretrain/lift/lift_pre_diffusion_mlp_ta4_td20/2024-06-28_14-47-58/checkpoint/state_5000.pt |
frozen DPPO diffusion policy, lift (DSRL release) | bcbaca212c40bf8632b039f6c81ff521 | 4,597,035 |
robomimic-pretrain/can/can_pre_diffusion_mlp_ta4_td20/2024-06-28_13-29-54/checkpoint/state_5000.pt |
frozen DPPO diffusion policy, can (DSRL release) | 5e418f9e590b220ad1a26a8dbab6f04c | 4,613,419 |
robomimic-pretrain/square/square_pre_diffusion_mlp_ta4_td100_ddim-100steps/2025-04-11_19-13-26_44/checkpoint/state_3000.pt |
frozen DPPO diffusion policy, square (DSRL release) | 412f245650f930842c12514f9e8fd49a | 18,454,562 |
robomimic/{lift,can,square}/normalization.npz |
observation/action normalization (DSRL release) | c09ff53a… / a04b2abe… / 2871f3f5… | 1,434 / 1,498 / 1,498 |
experts/lift/expert_dppo_td20.pt |
teacher, DPPO official fine-tune recipe (SR ≈ 0.995) | 41fc27e8806bdfc1179bc0fc14d0fabb | 5,150,298 |
experts/can/expert_dppo_td20.pt |
teacher, DPPO official fine-tune recipe (SR ≈ 1.00) | 86e2252878473524c057d53fabbb0eb4 | 5,170,814 |
experts/square/expert_dsrlfork.pt |
teacher, dsrl-fork recipe fine-tuned FROM the frozen td100 base (SR ≈ 0.94 under our protocol). Use this one for square — the td20-recipe square teacher is schedule-mismatched with the frozen sampler. | 3a4d08ce082fb3f9e87daca46e041ab0 | 19,012,808 |
The diffusion-policy checkpoints and normalization files are the DSRL authors'
public release (Google Drive tree, dppo/log layout, see
https://github.com/ajwagen/dsrl), mirrored here unchanged for reproducibility.
Teachers were produced with the DPPO fine-tuning recipes described in
docs/three_method_robomimic_results.md of the VAM repo.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support