feat: add motion imitation tracking for G1 humanoid

Migrate motion tracking implementation from mjlab to UniLab architecture.
Adapts BeyondMimic-style motion imitation to UniLab's non-manager based
environment system with CPU simulation and GPU training.

Key components:
- Math utilities for quaternion operations and transforms
- Motion loader with adaptive/uniform/start sampling modes
- G1 motion tracking environment with exponential reward functions
- MuJoCo backend extensions for body state access
- CSV to NPZ motion preprocessing script

Architecture differences from mjlab:
- Direct NpEnv inheritance vs manager-based system
- CPU MuJoCo simulation vs GPU MuJoCo-Warp
- Decorator-based registration vs automatic config registration
- Inline reward/observation computation vs manager terms

Training config: 4096 envs, 30k iterations, 50 Hz control

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: caozx1110

scripts/motion/README.md

@@ -0,0 +1,68 @@
+# Motion Preprocessing
+
+This directory contains scripts for preprocessing motion data for motion tracking tasks.
+
+## CSV to NPZ Conversion
+
+The `csv_to_npz.py` script converts motion data from CSV format to NPZ format with precomputed forward kinematics.
+
+### Input Format
+
+CSV files should contain motion data in Unitree's generalized coordinate convention:
+- Columns 0-2: Base position (x, y, z)
+- Columns 3-6: Base quaternion (x, y, z, w) - will be converted to wxyz internally
+- Columns 7+: Joint angles (29 joints for G1)
+
+### Output Format
+
+NPZ files contain:
+- `fps`: Frame rate (integer)
+- `joint_pos`: Joint positions (N_frames × N_joints)
+- `joint_vel`: Joint velocities (N_frames × N_joints)
+- `body_pos_w`: Body positions in world frame (N_frames × N_bodies × 3)
+- `body_quat_w`: Body quaternions in world frame (N_frames × N_bodies × 4, wxyz)
+- `body_lin_vel_w`: Body linear velocities (N_frames × N_bodies × 3)
+- `body_ang_vel_w`: Body angular velocities (N_frames × N_bodies × 3)
+
+### Usage
+
+```bash
+# Basic usage
+uv run python scripts/motion/csv_to_npz.py \
+  --input_file path/to/motion.csv \
+  --output_file path/to/motion.npz \
+  --input_fps 30 \
+  --output_fps 50
+
+# With custom model file
+uv run python scripts/motion/csv_to_npz.py \
+  --input_file path/to/motion.csv \
+  --output_file path/to/motion.npz \
+  --input_fps 30 \
+  --output_fps 50 \
+  --model_file path/to/model.xml
+
+# Process specific line range
+uv run python scripts/motion/csv_to_npz.py \
+  --input_file path/to/motion.csv \
+  --output_file path/to/motion.npz \
+  --input_fps 30 \
+  --output_fps 50 \
+  --line_range 100 500
+```
+
+### Parameters
+
+- `--input_file`: Path to input CSV file (required)
+- `--output_file`: Path to output NPZ file (required)
+- `--input_fps`: Frame rate of input CSV (default: 30)
+- `--output_fps`: Desired output frame rate (default: 50)
+- `--model_file`: MuJoCo model file (default: G1 flat scene)
+- `--line_range`: Line range to process [start, end] (optional)
+
+### Notes
+
+- The script uses LERP for position interpolation and SLERP for quaternion interpolation
+- Velocities are computed using numerical differentiation
+- Forward kinematics is computed using MuJoCo for all bodies
+- The output FPS should match the control frequency of your training environment (typically 50 Hz)