Downloads · 30 days
0
Myungkyu/hiwrld-baseline-ckpts-real-robot
hiwrld-baseline-ckpts-real-robot is a robotics model from Myungkyu. Use it for the robotics task on the model card, and read the license before you ship it in a product. It is set up for lerobot. The card lists the license as apache-2.0.
Low-level visuomotor policies fine-tuned on the two real-robot teleoperation datasets Myungkyu/instanceretrieval and Myungkyu/layoutreconstruction (80 episodes each, 10 Hz, exterior + wrist camera at 448×252, 8-D join…
Downloads · 30 days
0
Access
Public
Updated Sep 12, 2026
Repo size
584 GB
Likes
1
Public
Click a slice to open those files.
.safetensors584 GB · 100%
From the Hugging Face model README
Low-level visuomotor policies fine-tuned on the two real-robot teleoperation datasets
Myungkyu/instance_retrieval and
Myungkyu/layout_reconstruction
(80 episodes each, 10 Hz, exterior + wrist camera at 448×252, 8-D joint state).
Directory nomenclature: {pi0_5, rldx_1, rldx_1_img}_{instance_retrieval, layout_reconstruction}_{joint, eef}_b64_60k/step_XXXXXX/
| family | model | video input | action space | tasks | steps kept |
|---|---|---|---|---|---|
pi0_5_*_joint | π0.5 (lerobot/pi05_base init) | current frame | 8-D absolute joint targets + gripper | both | 010000–060000 |
pi0_5_*_eef | π0.5 | current frame | 7-D end-effector velocity + gripper | both | 010000–060000 |
rldx_1_img_*_joint | RLDX-1 (RLWRLD/RLDX-1-PT-IMG init) | current frame (video_length 1) | 8-D absolute joint targets + gripper | both | 010000–060000 |
rldx_1_img_*_eef | RLDX-1 (RLWRLD/RLDX-1-PT-IMG init) | current frame | 7-D end-effector velocity + gripper | both | 060000 |
rldx_1_*_eef | RLDX-1 (RLWRLD/RLDX-1-PT init) | 4-frame history (video_length 4) | 7-D end-effector velocity + gripper | both | 010000–060000 |
All runs: batch size 64, 60k optimisation steps, single-task fine-tuning on the corresponding dataset layout
(lerobot_v30_abs_joint / lerobot_v30_delta_eef for π0.5, lerobot_v21_abs_joint / lerobot_v21_delta_eef for RLDX-1).
The state input is the same 8-D vector (7 measured joint angles + gripper closedness) in every run; only the action head differs.
Language input: the dataset's task instruction.
config.json, model.safetensors, policy_preprocessor*.{json,safetensors}, policy_postprocessor*.{json,safetensors},
train_config.json — load with lerobot ≥ 0.5 (PI05Policy.from_pretrained(<dir>) + make_pre_post_processors(cfg, pretrained_path=<dir>));
state/action normalisation (quantiles) lives in the processors, so raw state goes in and actions come out in dataset units.config.json, model-0000X-of-00003.safetensors, model.safetensors.index.json, processor/ (modality config, normalisation
statistics, embodiment id), experiment_cfg/ (training configuration) — load with the RLDX-1 code base (RLDXPolicy(model_path=<dir>, ...)).
config.json model_name refers to the hub backbone RLWRLD/RLDX-1-VLM; the RLDX-1 loader resolves it through the Hugging Face cache.huggingface-cli download Myungkyu/hiwrld-baseline-ckpts-real-robot --include "pi0_5_instance_retrieval_joint_b64_60k/step_060000/*" --local-dir ckpt
huggingface-cli download Myungkyu/hiwrld-baseline-ckpts-real-robot --include "rldx_1_img_layout_reconstruction_joint_b64_60k/step_060000/*" --local-dir ckpt