pi05_real_v1 — EBiM Task 2 (thermal pad placement), real Mobile FR3 Duo

lerobot/pi05_base fully fine-tuned on the organizers' real-robot release ebim-benchmark/ebim_task2_realrobotdata (task2_munich), keeping the 149 reviewer-accepted episodes (annotation.human.validity == 1; 86,444 frames at 20 Hz). Primary Phase II weights: 4×H800, global batch 32, 30,000 steps (the openpi pi0.5 full-fine-tune recipe: AdamW, warmup 1k, cosine 2.5e-5 → 2.5e-6), final checkpoint. The single-GPU alternate is ByteMelodist/pi05_real_v1_dbg.

Observation / action contract

key	shape	source topic (LABS station names)
observation.images.base_0_rgb	720×1280×3	/head_camera/zed_node/rgb/color/rect/image
observation.images.left_wrist_0_rgb	480×640×3	/wrist_camera_left/camera/color/image_raw
observation.images.right_wrist_0_rgb	480×640×3	/wrist_camera_right/camera/color/image_raw
observation.state	8	right measured_joint_states joints 1–7 + right Robotiq knuckle (rad)
action	8	right joint targets 1–7 (rad) → /right/follower/gello/joint_states; gripper width percent (1 = open) → .../target_gripper_width_percent
task	text	Pick up the thermal pad and place it on the target RAM board

Spine fixed at 0.434 m (the dataset's constant 434 mm); left arm not commanded by the policy. Runner and Dockerfiles: https://github.com/cgboy520/ebim-task2-autonomy (Phase II section).

Loading

from lerobot.configs.policies import PreTrainedConfig
from lerobot.policies.factory import get_policy_class, make_pre_post_processors
cfg = PreTrainedConfig.from_pretrained("ByteMelodist/pi05_real_v1")
policy = get_policy_class(cfg.type).from_pretrained("ByteMelodist/pi05_real_v1")
pre, post = make_pre_post_processors(cfg, pretrained_path="ByteMelodist/pi05_real_v1")

lerobot 0.6.0, torch ≥ 2.7.

Tokenizer

tokenizer/ is a copy of the PaliGemma text tokenizer (google/paligemma-3b-pt-224), included so the policy loads without Hub access. Like the weights above it is a derivative of Gemma: Gemma is provided under and subject to the Gemma Terms of Use.