mertalbaba/Robo-Xperience10M-GMR

Robo-Xperience-10M GMR

Per-episode GMR (Generalized Motion Retargeter) outputs mapping human SMPL motion from the Xperience-10M egocentric dataset onto the Unitree G1 unitree_g1_with_hands URDF. One motion.pkl per episode, preserving the original <uuid>/<ep>/ directory structure. 12,492 episodes.

Layout

<uuid>/<ep>/motion.pkl

Each motion.pkl is a Python pickle containing a dict with keys:

key	shape / type	notes
fps	int64 scalar	Source human-capture rate. Varies per episode (20–25 Hz observed). Do NOT assume a single rate downstream.
root_pos	(T, 3) float64	Pelvis position in world frame (world Z is up).
root_rot	(T, 4) float64	XYZW quaternion (scipy convention), NOT WXYZ. See ⚠ below.
dof_pos	(T, 43) float64	Joint positions, MuJoCo order for unitree_g1_with_hands. See ⚠ below.
local_body_pos	object	Per-link body positions in local frame.
link_body_list	object	Names of the bodies indexed by local_body_pos.
source_hdf5	str	Path to the source Xperience annotation.hdf5.
robot	str	Always "unitree_g1_with_hands" for this dataset.
actual_human_height	float	Source subject's measured height (metres).

⚠ Consumer caveats

1. root_rot is XYZW, not WXYZ

GMR's retargeter internally uses MuJoCo's WXYZ (scalar-first) qpos[3:7] convention, but the save-time code (scripts/smplx_to_robot.py:154) explicitly reorders to XYZW (scipy .as_quat() default) before pickling:

root_rot = np.array([qpos[3:7][[1, 2, 3, 0]] for qpos in qpos_list])

Empirical evidence that on-disk is XYZW: pick any episode's root_rot and compute the rotation angle assuming each convention. For a mostly-upright walking human the median tilt should be ~10–30°, not ~170°:

import numpy as np, pickle
q = pickle.load(open("<uuid>/<ep>/motion.pkl", "rb"))["root_rot"]
q = q / np.linalg.norm(q, axis=1, keepdims=True)
theta_xyzw = 2 * np.arccos(np.clip(np.abs(q[:, 3]), 0, 1)) * 180 / np.pi
theta_wxyz = 2 * np.arccos(np.clip(np.abs(q[:, 0]), 0, 1)) * 180 / np.pi
print(f"median rotation — XYZW: {np.median(theta_xyzw):.1f}°  WXYZ: {np.median(theta_wxyz):.1f}°")
# Sample episode returned: XYZW=26.8°  WXYZ=176.0°  → XYZW is correct.

If you feed root_rot directly to MuJoCo's data.qpos[3:7] (which expects WXYZ), you MUST reorder first:

root_rot_wxyz = root_rot_xyzw[[3, 0, 1, 2]]     # -> WXYZ for MuJoCo

Any helper that assumes WXYZ (e.g. quaternion-scalar-first math like q[0] = w) will silently produce wrong rotations if fed the raw disk value.

2. dof_pos layout and empty finger slots

dof_pos is 43-dim in MuJoCo order for the unitree_g1_with_hands URDF:

slice	joints
[0:22]	body: left leg (6) + right leg (6) + waist (3) + left arm (7)
[22:29]	left hand (7) — always zero (GMR does not fit fingers)
[29:36]	right arm (7)
[36:43]	right hand (7) — always zero (GMR does not fit fingers)

If you need body-only 29-dim MJ-order (legs + waist + both arms), take dof_pos[:, [0:22]] + dof_pos[:, [29:36]] — concatenate the two non-finger blocks. Hand motion for this dataset is provided separately by a MANO-based tokenizer (see the sibling Robo-Xperience10M-Hand dataset).

3. Variable source fps

fps is per-episode and reflects the source human capture rate. Both 20 Hz and 25 Hz have been observed; some episodes have fractional rates because the source length_sec is not an integer number of frames divided by nominal fps. Downstream code should read fps from the pkl, not hardcode it.

4. World frame convention

root_pos is world-frame; on the sampled episode root Z ∈ [0.43, 1.03] m — hip height of a walking human. World Z is up. Xperience source is Y-up SMPL; GMR's retargeter strips the SMPL Y-up basis before writing.

Provenance

• Source: Xperience-10M (SMPL egocentric mocap, 12.7k episodes).
• Retargeter: GMR — General Motion Retargeter.
• Target embodiment: Unitree G1 with hands (43 DOF URDF).