A Practical 2026 Guide for Robotics Engineers
Training a humanoid robot to run stably is significantly more difficult than teaching it to walk. Running involves higher speeds, stronger impact forces, dynamic balance, and precise foot placement. The quality and structure of your locomotion pose dataset can make or break your training results.
In this guide, we’ll walk through exactly how to use high-quality pose estimation data (such as QualityVision’s Jogging and Running packs) to train stable and natural running behavior in humanoid robots.
1. Why Standard Pose Data Often Fails for Running
Most public datasets (Human3.6M, AMASS, 3DPW) have limitations when used for running:
Lack of temporal smoothing → jittery keypoints
No body normalization → inconsistent skeleton scale
Missing rich motion features (velocity, acceleration, stride)
Low motion consistency in dynamic movements
This leads to unstable gaits, foot sliding, and poor balance in simulation and real robots.
2. Recommended Dataset Requirements for Humanoid Running
For successful running training, your dataset should include:
High mean quality score (> 0.85 recommended)
Gaussian temporal smoothing to reduce jitter
Hip-center body normalization for consistent skeleton scaling
Rich kinematic features: velocity, acceleration, stride length, motion consistency
Diverse real-world videos covering different speeds and environments
Strict quality filtering with transparent reports
QualityVision Jogging Pack (61 videos – 14,550 HQ frames, mean quality 0.857) and Running Pack (106 videos – 25,944 frames) were specifically processed to meet these requirements.
3. Step-by-Step: How to Train Stable Running
Step 1: Data Preparation
Download the full JSONL files (data.jsonl + per_video/ splits)
Use the features.json for sequence-level metrics (velocity, stride, motion_consistency)
Apply the provided keypoints_body_normalized field (hip-center + torso scale) — this is critical for policy stability
Step 2: Feature Engineering
Extract or use directly:
Hip-center velocity (x, y, z)
Stride length and step frequency
Motion consistency score
Lower-body visibility and joint angles
These features serve as excellent reward signals or state observations in reinforcement learning.
Step 3: Training Pipeline (Recommended)
Imitation Learning (Phase 1)
Use Behavioral Cloning or GAIL on the smoothed normalized keypoints
Focus on hip trajectory and foot placement
Reinforcement Learning (Phase 2)
Combine pose data with physics simulation (Isaac Gym, MuJoCo, ORBIT, etc.)
Reward function ideas:
Forward velocity matching
Motion consistency penalty
Foot contact stability
Energy efficiency
Domain Randomization
Add noise to the pose data during training
Use the built-in augmentations (horizontal flip + keypoint noise) from the dataset
Step 4: Fine-tuning on Your Robot
Transfer the policy from simulation to real hardware
Use the high-quality jogging/running data to reduce the sim-to-real gap
4. Pro Tips from Real Experiments
Start with Jogging data before full-speed Running — it provides better stability for initial policy learning.
Prioritize lower-body visibility and motion_local metrics when filtering frames.
Use the gaussian_smoothed coordinates (window=5) instead of raw keypoints for smoother reward signals.
Track motion_consistency during training — a drop below 0.5 usually indicates unstable running.
5. Ready-to-Use Datasets for This Workflow
QualityVision Jogging Pose Dataset — 14,550 HQ frames, ideal starting point for stable running
QualityVision Running Pose Dataset — 25,944 frames for high-speed training
Full Locomotion Bundle — Complete coverage (Walking + Jogging + Running)
All packs include:
Clean JSONL format
Temporal smoothing applied
Body normalization
Full quality reports and rejected frames
Browse and purchase the datasets here →