metadata
library_name: pytorch
tags:
- robotics
- diffusion-policy
- ddpm
- manipulation
license: apache-2.0
ARC-AI Diffusion Policy
Denoising Diffusion Probabilistic Model trained for 7-DOF robotic manipulation.
Specs
| Parameter | Value |
|---|---|
| Parameters | 1,503,751 |
| Architecture | DDPM + Temporal Encoder |
| Training Steps | 500,000 |
| Best Loss | 0.04596 |
| Action Horizon | 16 steps |
| Denoising (inference) | 10 steps (DDIM) |
| Observation Dim | 20 |
| Action Dim | 7 |
Usage
import torch
model = torch.load("checkpoints/diffusion_policy_500k.pt")
Training
Trained on 10K expert demonstrations (Jacobian IK, MuJoCo Franka) on NVIDIA A100-SXM4-80GB.