Dongkkka/evo1_dashboard_0904_5k_8bs

Evo1 — final 5,000-step checkpoint

Batch 8, action chunk 16, three cameras, 22-dimensional state and actions. Trained with LeRobot 0.6.1 in Stage 1: the pretrained InternVL3 backbone is frozen, and the newly initialized action head is trained. The policy contains approximately 747 million parameters, with 94 million trainable parameters during this run.

Full policy weights and the saved preprocessing and normalization are included. The first 19 action channels are absolute position commands and the last three are base velocities. Gripper binarization and relative-action conversion are disabled. Robot task success has not been evaluated.

Standard LeRobot loading

from lerobot.policies.evo1.modeling_evo1 import Evo1Policy
from lerobot.policies.factory import make_pre_post_processors

repo_id = "Dongkkka/evo1_dashboard_0904_5k_8bs"
policy = Evo1Policy.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 its Evo1 dependencies. No custom loader is required. The standard implementation may download OpenGVLab/InternVL3-1B-hf during initialization; its assets must be available or cached. The training run used base revision 014c0583a0d4bedf29fbe2dbff4f865eb998e171. Provide all three camera keys from config.json, the robot state, and task text. Use the saved preprocessor and postprocessor to preserve action normalization.