Research

Train a Humanoid from AI Video, Not MoCap: NCKU’s Synthetic Demo Pipeline

Robb Harlan 4 min read

Collecting real humanoid demonstrations is expensive, narrow, and weirdly personal—every person does the same task differently. A National Cheng Kung University (NCKU) team proposes skipping MoCap entirely: generate diverse human motion videos from text prompts, retarget them to a Unitree G1, and train RL motion-tracking policies in simulation.

The paper, Learning Diverse Humanoid Tasks via Synthetic Video Scenarios without Real World Data (arXiv:2607.21648, 22 July 2026), is pure sim for now—but the pipeline is concrete.

Pipeline diagram from prompts through Veo 3 to humanoid RL training
Framework overview: prompts → generative video → pose retargeting → RL tracking. Source: arXiv HTML (2607.21648).

How the pipeline works

Authors Yun-Hao Tsai, Cong-Thanh Vu, and Yen-Chen Liu describe three stages:

  1. Prompt-driven video generation — structured instruction prompts plus user task prompts drive Google Veo 3 (and related generative models) to produce full-body, camera-stable human motion clips. Multiple videos per task capture different execution styles.
  2. Video → humanoid reference — SMPL-X style pose extraction, then retargeting to the humanoid (GMR) with limb scaling, foot-slide fixes, and IK under joint limits.
  3. Motion stitching + RL tracking — root alignment and joint smoothing to chain clips; DeepMimic-style RL (PPO in NVIDIA Isaac Lab) with imitation rewards on pose, velocity, end-effectors, and CoM, plus limit/smoothness/self-collision penalties.

They report training with 4096 parallel agents on Isaac Lab hardware (Xeon W5-3435X + RTX 4000 Ada in their setup).

Side-by-side AI video frames and Unitree G1 mimic for lie-stand, boxing, pick-and-place
Qualitative results: generative human video vs G1 mimic on four task families. Source: arXiv HTML (2607.21648).

What they evaluated

  • Motion generation: 50 everyday task prompts × 10 videos each; most everyday motions looked physically plausible; high-speed actions (e.g. backflips) still showed temporal glitches.
  • Policy tasks: lie-and-stand, boxing, pick-and-place (including with a 0.5 kg payload). Joint-position MAE roughly 0.04–0.07 m on dynamic tasks; loads hit the upper body harder while lower-body tracking stayed steadier.

The conclusion is careful: the method works in simulation for diverse whole-body tasks; sim-to-real and physical consistency of generated video are left for future work.

A Human’s Take

If generative video can flood the imitation dataset with style variation, humanoid learning stops waiting for a motion-capture studio booking. That is a real data bottleneck unlock—on paper.

I care about two follow-ups: (1) does this survive domain gap to a real G1 without collapsing, and (2) how often Veo invents physics that the retargeter happily teaches the robot. Diversity is worthless if half the demos are soft-body fiction.

Sources