Atarilab/diplo-dataset

DiPLO Pick-and-Place Dataset (Unitree G1)

The training dataset of DiPLO: Self-distilled Diffusion Planning for Loco-manipulation (arXiv:2610.24847, project page).

It contains whole-body humanoid box pick-and-place motions for the Unitree G1 (29 DoF). Each motion comes as a pair:

• a kinematic reference: human motion retargeted to the robot, taken from the OmniRetarget dataset. It can be physically imperfect (foot sliding, penetration, unbalanced contacts).
• a tube of physically simulated trajectories: several sampling-based trajectory optimization (SBTO) solves in MuJoCo that track that reference.

These are the raw motions, before any training-time preprocessing. All augmentations used for the final DiPLO models are already included in the files: the D4 box symmetry and the SBTO tube.

At a glance

Robot	Unitree G1, 29 DoF (g1_29dof_rev_1_0) + one free-floating box
Source motions	29 largebox pick-and-place clips, 5 subjects (sub3, sub7, sub10, sub12, sub16)
Symmetry augmentation	8 D4 variants per clip (identity, rot90, rot180, rot270, each ± mirror) → 232 directories
SBTO augmentation	4–39 solves per directory (median 36) → 7184 trajectories
Frame rate	50 Hz
Length	294–572 frames (5.9–11.4 s)
Size	~6.3 GB, float32 NumPy .npz

Whole-body examples: six source clips, each replaying one of its SBTO trajectories on the robot and box, with the base path drawn on the floor.

Top-down replay of the 29 source clips (left) and their mirror variants (right). Every SBTO trajectory draws its robot-base (pelvis) path in cyan, the kinematic reference is the pale line, and the box's footprint is shown at its true size and yaw. The D4 rotation variants turn the box about its own axis and leave the base path essentially unchanged, so they are not shown separately.

Layout

<motion>[__<augmentation>]/
├── kinematic_traj.npz   # (1, T, …) kinematic reference
└── sbto_traj.npz        # (B, T, …) SBTO tube, B solves
metadata/
└── conversion_meta.json # build settings, per-motion solver cost / kept indices

Directory names are the source motion name, plus a suffix for the symmetry variant: sub10_largebox_033_original, sub10_largebox_033_original__rot90_mirror, and so on. The un-augmented clips match ^sub[0-9]+_largebox_[0-9]+_original$.

Array schema

Both files share the same keys. The leading axis is the batch: 1 for the kinematic reference, B for the tube. Then comes time T. All quantities are in the world frame, both linear and angular velocities. Quaternions are wxyz.

Key	Shape	Description
root_pos, root_rot	(·, T, 3), (·, T, 4)	pelvis position / orientation
root_lin_vel, root_ang_vel	(·, T, 3)	pelvis velocities
dof_pos, dof_vel	(·, T, 29)	joint positions / velocities
dof_pd_target	(·, T, 29)	PD position targets
joint_pos, joint_vel, joint_pd_targets	(·, T, 29)	aliases of the three above
body_pos_w, body_quat_w	(·, T, 30, 3/4)	all 30 body poses
body_lin_vel_w, body_ang_vel_w	(·, T, 30, 3)	all 30 body velocities
object_pos, object_rot	(·, T, 3/4)	box pose
object_lin_vel, object_ang_vel	(·, T, 3)	box velocities
object_*_w		aliases of the object arrays
contact_indicators	(·, T, 4) bool	contact flags for [left_foot, right_foot, left_hand, right_hand]: feet vs. floor, hands vs. box
fps	(1,)	50.0

Joint order (MuJoCo order): left_hip_pitch, left_hip_roll, left_hip_yaw, left_knee, left_ankle_pitch, left_ankle_roll, right_hip_pitch, right_hip_roll, right_hip_yaw, right_knee, right_ankle_pitch, right_ankle_roll, waist_yaw, waist_roll, waist_pitch, left_shoulder_pitch, left_shoulder_roll, left_shoulder_yaw, left_elbow, left_wrist_roll, left_wrist_pitch, left_wrist_yaw, right_shoulder_pitch, right_shoulder_roll, right_shoulder_yaw, right_elbow, right_wrist_roll, right_wrist_pitch, right_wrist_yaw (each with the suffix _joint).

Body order: pelvis, then the *_link of each joint above in the same order. torso_link takes the place of waist_pitch.

How it was built

1. Kinematic references. The 29 largebox clips are taken from the OmniRetarget dataset (paper): OMOMO human-object demonstrations retargeted to the G1. They are aligned to the origin, so that xy = 0 and yaw = 0 at frame 0.
2. SBTO tube. For each reference, a sampling-based trajectory optimizer solves in MuJoCo for physically consistent trajectories that track it.
3. Symmetry. Every clip, including both the reference and its tube, is expanded over the 8 elements of the box's D4 symmetry group (yaw rotations by 90° and mirroring).

Known limitations

• Raw solver output. The SBTO solves are not curated, and some may not be executable by a feedback controller.
• Tail interpolation can penetrate the box. At the end of each clip we append an interpolation from the reference's final pose to the robot's default joint positions. This tail is not simulated, and the arms or legs sometimes pass into the box during it. So far this has not noticeably affected our training.
• Small object set. Only one box geometry (largebox) and top grasps.

Usage

hf download Atarilab/diplo-dataset --repo-type dataset --local-dir diplo-dataset

import numpy as np

d = "diplo-dataset/sub10_largebox_033_original"
ref = np.load(f"{d}/kinematic_traj.npz")   # ref["dof_pos"].shape == (1, T, 29)
tube = np.load(f"{d}/sbto_traj.npz")       # tube["dof_pos"].shape == (B, T, 29)

Citation

@misc{omar2026diplo,
  title         = {{DiPLO}: Self-distilled Diffusion Planning for Loco-manipulation},
  author        = {Omar, Shafeef and Sanokowski, Sebastian and Daneshmand, Elham and
                   Yu, Dian and Cremers, Daniel and Schmid, Cordelia and Khadiv, Majid},
  year          = {2026},
  eprint        = {2610.24847},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2610.24847}
}