XYZPIT/pick_up_the_specified_breadbag_rgb

Pick up the specified bread bag — RGB baseline, H16, 40k

GR00T-N1.6-3B RGB baseline fine-tuned for bread bag picking with both arms and grippers. This repository contains only checkpoint-40000 from rgb_baseline_h16_20261023.

• Action horizon: 16.
• Inputs: head RGB, left wrist RGB, right wrist RGB, robot state, and task instruction.
• Action outputs: both arms and grippers (14 dimensions).
• RGB inputs without blur or tint emphasis.
• Model weights, processor configuration, embodiment mapping, and normalization statistics are included.

Instructions

• Pick up the leftmost bread bag with the left hand.
• Pick up the leftmost bread bag with the right hand.
• Pick up the middle bread bag with the left hand.
• Pick up the middle bread bag with the right hand.
• Pick up the rightmost bread bag with the left hand.
• Pick up the rightmost bread bag with the right hand.

Download and use

from huggingface_hub import snapshot_download
from gr00t.data.embodiment_tags import EmbodimentTag
from gr00t.policy.gr00t_policy import Gr00tPolicy

model_path = snapshot_download(repo_id="XYZPIT/pick_up_the_specified_breadbag_rgb")
policy = Gr00tPolicy(
    embodiment_tag=EmbodimentTag.NEW_EMBODIMENT,
    model_path=model_path,
    device="cuda",
)

Use with the compatible Isaac-GR00T N1.6 environment. Retain all processor and statistics files when downloading.

Checkpoint selection

Selected for balanced arm tracking and gripper behavior across six recorded training episodes (13, 14, 102, 50, 10, 113), using open-loop evaluation with horizon 16 and seed 42 + episode ID. Mean arm MSE: 0.000216134; mean gripper MSE: 40.697402; overall native-unit MSE: 5.814100. These are equal averages across episodes, not held-out validation or closed-loop success rates.