PI0.5 Fine-tuned on Droyd UMI Data
Fine-tuned PI0.5 on bimanual manipulation demonstrations collected via UMI (Universal Manipulation Interface) gloves.
Training Details
| Parameter |
Value |
| Base model |
lerobot/pi05_base (4.1B params) |
| Dataset |
110 approved episodes, 11,765 frames @ 50fps |
| Session |
158cc246-6594-410a-8d5a-db58a4101276 |
| Steps |
5,000 |
| Batch size |
32 effective (4 per GPU × 8 GPUs) |
| Hardware |
8× NVIDIA H200 (Vast.ai) |
| Training time |
~62 minutes |
| Final loss |
~0.002 |
| Optimizer |
AdamW (lr=2.5e-5, weight_decay=0.01) |
| LR schedule |
Cosine decay with 166-step warmup |
| Normalization |
Quantile (state & action) |
| Action space |
7D absolute joint positions (6 joints + gripper) |
| Cameras |
ego (overhead) + wrist |
| Review clips |
Enabled (user-trimmed episode boundaries) |
Input Features
| Feature |
Shape |
Type |
observation.images.cam_high → base_0_rgb |
(3, 224, 224) |
VISUAL |
observation.images.cam_wrist → left_wrist_0_rgb |
(3, 224, 224) |
VISUAL |
observation.state |
(32,) |
STATE (quantile normalized) |
Output Features
| Feature |
Shape |
Type |
action |
(7,) |
ACTION (quantile normalized) |
W&B Run
droyd/lerobot/rmr4moga
Usage
from lerobot.common.policies.pi05.modeling_pi05 import PI05Policy
policy = PI05Policy.from_pretrained("rayhanfahmed/pi05_droyd_umi")
