EmbodyX/pi05_unitree_load_bottle_water_10k

π₀.₅ — Unitree G1 "Load the bottle water to the shelf" (step 10,000)

Fine-tuned π₀.₅ checkpoint for the Unitree G1 (Dex1 bimanual, waist-inclusive) trained with openpi.

This is an intermediate checkpoint taken at step 10,000 of a 30,000-step run.

Task

Language instruction (must be character-identical at inference):

Load the bottle water to the shelf

Action / state space — 17 dims

Unlike the standard openpi Unitree G1 configs (16 dims), this model includes the waist yaw joint:

index	joint
0–6	left arm (ShoulderPitch, ShoulderRoll, ShoulderYaw, Elbow, WristRoll, WristPitch, WristYaw)
7–13	right arm (same order)
14	left gripper
15	right gripper
16	waist yaw

Actions are absolute joint positions (no delta transform). Model outputs are padded to 32 dims and sliced back to the first 17.

Cameras

Three RGB streams, all mapped to real (unmasked) model slots:

dataset key	model slot
cam_left_high	base_0_rgb
cam_left_wrist	left_wrist_0_rgb
cam_right_wrist	right_wrist_0_rgb

Training

Base model	gs://openpi-assets/checkpoints/pi05_base/params
Dataset	XiaoweiLinXL/unitree_load_bottle_water (LeRobot v2.1, 132 episodes, 157,595 frames, 30 fps)
Config	pi05_unitree_g1_load_bottle_water
Action horizon	30 (~1 s at 30 fps)
Batch size	32
Optimizer	EMA decay 0.999
FSDP devices	2
Normalization	quantile (q01/q99)
Step	10,000 / 30,000

Loss at step 10,000: ~0.0055 (from 0.133 at step 0).

Contents

Only the inference-relevant parts of the openpi checkpoint are included:

params/                # model weights (12 GB)
assets/                # normalization statistics
_CHECKPOINT_METADATA

The train_state/ directory (optimizer state, ~31 GB) is not included, so this checkpoint can be used for inference/serving but not to resume training.

Usage

uv run scripts/serve_policy.py policy:checkpoint \
  --policy.config=pi05_unitree_g1_load_bottle_water \
  --policy.dir=/path/to/this/checkpoint

Requires the matching pi05_unitree_g1_load_bottle_water config (with action_dim=17) in your openpi checkout.