Dongkkka/eo1_dashboard_0904_4k_8bs

EO1: 4,000 steps, batch 8

Intermediate 4,000-step checkpoint from a 5,000-step run, action chunk 16, three cameras, 22D state/actions. All 28 dataset episodes used for training; no held-out validation split. The pretrained vision-language backbone was frozen; the newly initialized action head was trained. This is a full policy checkpoint with saved pre/post processors. The first 19 action channels are absolute position commands, followed by three base velocities. Gripper binarization and relative-action conversion are disabled. Robot task-success rate is not established by the loading check.

Standard LeRobot loading

from lerobot.policies.eo1.modeling_eo1 import EO1Policy
from lerobot.policies.factory import make_pre_post_processors

repo_id = "Dongkkka/eo1_dashboard_0904_4k_8bs"
policy = EO1Policy.from_pretrained(repo_id).to("cuda").eval()
preprocessor, postprocessor = make_pre_post_processors(
    policy_cfg=policy.config,
    pretrained_path=repo_id,
    preprocessor_overrides={"device_processor": {"device": "cuda"}},
)

Use LeRobot 0.6.1 with the policy dependencies. The base model/tokenizer assets may be downloaded during standard initialization. Training used base revision 66285546d2b821cf421d4f5eb2576359d3770cd3. Provide the three image keys in config.json, the 22D robot state, and task text. Always apply the saved preprocessor and postprocessor. No custom loading script is needed.