DiT Diffusion xArm Wire Rope v0
This repository contains a Multi-Task DiT diffusion policy fine-tuned with LeRobot on the JayCao99/xarm-wire-rope-v0 dataset.
Training Summary
| Field | Value |
|---|---|
| Policy | Multi-Task DiT |
| Objective | Diffusion |
| Dataset | JayCao99/xarm-wire-rope-v0 |
| Robot | xArm7 |
| Episodes | 210 |
| Frames | 290,153 |
| Cameras | ego, wrist |
| State dim | 15 |
| Action dim | 8 |
| Training steps | 100,000 |
| Batch size | 32 |
| Save frequency | 5,000 steps |
| GPU | 1x A100 80GB, Unity superpod-a100 |
| Runtime | approximately 8h 23m |
| Peak VRAM from log | approximately 21.72 GB |
| WandB | disabled |
Checkpoints
| Checkpoint | Notes |
|---|---|
checkpoint-100000/ |
Final policy checkpoint after 100k finetuning steps |
Usage
from huggingface_hub import snapshot_download
ckpt_dir = snapshot_download(
repo_id="Himanshu77275/dit-diffusion-xarm-wire-rope-v0",
allow_patterns="checkpoint-100000/*",
)
The checkpoint directory contains the LeRobot policy config, model weights, and dataset normalization/preprocessing files needed for policy loading.