The Unitree Myth Challenged: Embodied AI Still Has a Long Road Ahead
Opinion | August 27, 2026
Author: David Tang, Managing Director, Standard Kepler
Originally published in HKET (香港經濟日報)
Executive Summary
Unitree Technology (滬: 688836) made a spectacular debut on its IPO day, opening at RMB 1,100 — a 629% premium to its RMB 150.80 issue price. Early investors who won the lottery saw paper gains exceeding RMB 470,000 per lot. Founder Wang Xingxing’s net worth briefly touched RMB 130 billion. But beneath the hype lies a sobering reality: the path to genuine embodied intelligence remains far longer than the market assumes.
The IPO Phenomenon
Unitree’s market debut was nothing short of extraordinary:
| Metric | Value |
|---|---|
| Issue Price | RMB 150.80 |
| Opening Price | RMB 1,100 |
| First-Day Pop | 629% |
| Paper Gain (per lot) | > RMB 470,000 |
| Founder Net Worth (peak) | ~RMB 130 billion |
The company became a household name after its viral Spring Festival Gala performances — first the Yangge dance in 2025, then martial arts kung fu in 2026. From stiff, robotic movements to fluid acrobatics in just one year, the progress was undeniably impressive.
But here is the cold water: these movements, particularly leg locomotion, were built on foundations laid during the Boston Dynamics era. Jumping, running, falling and recovering — all these algorithms have been available in open-source communities for years. The Spring Festival spectacle was the harvesting of existing research, not a breakthrough.
What Is Embodied AI?
To understand humanoid robotics, we must first understand embodied intelligence (具身智能). Large language models can describe everything about an apple but cannot pick one up. Industrial robotic arms can perform precise operations but cannot understand a sentence. Embodied AI aims to bridge this gap: a robot with a brain that can, without pre-programmed instructions, turn a vague command into a sequence of physical actions while responding to real-world changes.
Three Components of Embodied AI
| Component | Description | Current State |
|---|---|---|
| Output (Action) | Physical movement: walking, grasping, manipulating | Legs: mature; Hands: nascent |
| Input (Perception) | Vision, tactile sensing, spatial awareness | Rapidly improving |
| Thinking (Planning) | Reasoning, task decomposition, adaptation | Early stage |
Legs vs. Hands: The Asymmetric Challenge
Legs: A Solved Problem (Mostly)
Leg locomotion is highly repetitive and predictable. The underlying algorithms were effectively solved by:
- ETH Zurich (2019): Legged locomotion control frameworks
- NVIDIA Isaac Gym (2021): Physics simulation for robot training
The difference between robots today is primarily hardware engineering, not software breakthroughs.
Hands: The Real Bottleneck
Hand manipulation is orders of magnitude more complex:
| Factor | Legs | Hands |
|---|---|---|
| Degrees of freedom | 6 (per leg) | 15+ (per hand) |
| Task periodicity | Highly repetitive | Non-periodic |
| Physical trajectory | Predictable | Unique per task |
| Interaction objects | Ground only | Unlimited variety |
| Complexity level | Pattern execution | Game-theoretic interaction |
Hand algorithms are a game against the universe of objects — each task has a unique physical trajectory.
Two Hardware Approaches
Tendon-Driven (腱驅動)
- Mechanism: High-strength cables replace gears and linkages, mimicking biological muscles and tendons
- Advantages: Inherent elasticity provides shock absorption; safer human interaction
- Disadvantages: Lower payload capacity; cable fatigue and precision degradation over time
Motor-Driven (電機驅動)
- Mechanism: Direct motor actuation with gears and linkages
- Advantages: Higher precision and payload capacity; predictable wear patterns
- Disadvantages: Rigid impact forces; less safe for human proximity
Both approaches are still actively explored. Neither has emerged as the definitive solution for general-purpose manipulation.
The Reality Check: Can It Hold a Coffee Cup?
Unitree’s robots can perform backflips and martial arts. But can they hold a coffee cup steadily? The gap between spectacular demonstrations and reliable utility is where the embodied AI field currently lives.
A robot doing a backflip is a circus act. A robot reliably making coffee is a product.
The challenge is not just hardware — it is the integration of perception, planning, and action in unstructured environments. A backflip requires precise control in a controlled setting. Making coffee requires understanding context, adapting to variation, and recovering from unexpected events.
Conclusion: Patience Required
Unitree’s IPO pop reflects investor enthusiasm for robotics, not the maturity of embodied AI. The field has made remarkable progress in locomotion, but manipulation — the true test of general-purpose utility — remains in its early stages.
For investors, the key takeaway is: hardware demonstrations are not product readiness. The path from viral video to reliable product is longer than the market currently prices.
Standard Kepler Research | standardkepler.com