Diffusion Policy — NutAssemblySquare (Square/PH/low-dim)
A robomimic DiffusionPolicyUNet checkpoint (UNet + DDPM noise scheduler,
action-chunked receding-horizon control: observation_horizon=2,
action_horizon=8, prediction_horizon=16) trained for 2000 epochs on
robomimic's public 200-demo Square/PH/low-dim dataset.
Rollout success rate (20 episodes, evaluated every 200 epochs):
| Epoch | 200 | 400 | 600 | 800 | 1000 | 1200 | 1400 | 1600 | 1800 | 2000 |
|---|---|---|---|---|---|---|---|---|---|---|
| Success | 85% | 85% | 85% | 70% | 75% | 90% | 85% | 85% | 95% | 90% |
checkpoints/model_epoch_1800_low_dim_success_0.95.pth (95% success, first
epoch to hit the run's peak) is the checkpoint used as the frozen base
policy for downstream residual-RL experiments.
This is a second, parallel base-policy lineage alongside a BC-RNN baseline trained on the same task/dataset, for a residual-RL project comparing how a heuristic-triggered SAC residual correction interacts with each base policy.
Downloading a specific checkpoint:
from huggingface_hub import hf_hub_download
hf_hub_download(
"georginio2000/diffusion-square-nutassembly",
"checkpoints/model_epoch_1800_low_dim_success_0.95.pth",
)