Figure 02 and the Humanoid Robot Race: Who Wins?
The humanoid robot race is no longer science fiction: it is an industrial war where billions of dollars clash to define the future of physical work. Between the mechanical elegance of Figure 02, the financial firepower of Tesla Optimus, the martial agility of Unitree G1, and the domestic approach of 1X Neo, which model will dominate the market? In this article, we will dissect the technical capabilities, pricing strategies, behind-the-scenes investors, and commercialization dates of these four contenders for the throne to understand who truly has the edge.
Prerequisites
- Understand the basics of Reinforcement Learning applied to robotics
- Have a grasp of actuators and robot body dynamics
- Know the differences between LLM (Large Language Models) and VLM (Vision-Language Models) architectures
- Be familiar with the concepts of Edge inference (embedded computing) vs Cloud
- Understand what "sim-to-real transfer" (transferring skills from a simulator to the real world) is
The Context: Why 2024-2025 is the Tipping Point
According to recent analyses by The Verge, the year 2024 marks a paradigm shift. Until now, humanoid robots were slow, wired prototypes incapable of adapting to unstructured environments. The convergence of three factors changed everything: the explosion of multimodal AI models, the drastic drop in actuator costs, and the massive influx of venture capital. The market, once dominated by Boston Dynamics with solutions priced at PH_prix_boston_dynamics_PH dollars, is seeing startups emerge that are ready to break the PH_prix_cible_commercial_PH dollar barrier to democratize access.
Figure 02: The High-Tech Challenger Backed by Giants
Figure AI is probably the company that generated the most buzz in 2024 with the release of its second model, the Figure 02. Unlike other players who tinker with existing components, Figure AI adopted a "vertically integrated" approach by designing its own actuators, its own battery, and its own computing architecture.
Figure 02 Technical Specifications
The Figure 02 is not just a simple update; it is a generational leap. It directly integrates a multimodal language model (VLM) developed in partnership with OpenAI, allowing it to reason about what it sees and hears in real time.
# Figure 02 spec sheet
figure_02:
apparence:
hauteur: "1.70 m"
poids: "70 kg"
charge_utile: "20 kg"
materiel:
structure: "Aluminium and carbon fiber composite"
mains: "16 degrees of freedom, tactile force sensors"
informatique:
gpu_embarque: "Nvidia Jetson Orin (x3)"
ia: "Custom OpenAI VLM (visual and vocal reasoning)"
cameras: "6 RGB cameras"
autonomie:
batterie: "2.25 kWh (designed in-house)"
duree: "5 hours of standard operation"
capacites_cles:
- "Bidirectional natural conversation"
- "Real-time visuo-motor correlation"
- "Manipulation of non-standardized objects"
Figure AI Investors and Strategy
Figure AI raised a staggering amount of {{levée_fonds_figure}} million dollars. Their strength lies in their board of directors and investors: Jeff Bezos, Nvidia, Microsoft, and OpenAI. This alliance gives them privileged access to the most powerful chips and the most advanced AI models before their competitors.
Figure AI's targeted pricing strategy is aggressive. While the Figure 01 was estimated at over $100,000, the industrial goal for the Figure 02 is to drop below the PH_prix_cible_figure_PH dollar mark at mass production scale, primarily for logistics and automotive assembly.
Tesla Optimus: The Advantage of the Industrial Monster
Tesla is not playing in the same league. While Figure AI has to build its factory, Tesla already possesses the manufacturing infrastructure (Giga factories) and the software ecosystem (Full Self-Driving) to accelerate its development.
The "Minimalist Hardware, Maximalist Software" Approach
The Tesla Optimus (Gen 2) bets on an exponential reduction in weight and extreme cost optimization. Elon Musk stated that large-scale production could begin as early as 2025, with a mind-boggling target price of under $20,000. However, public demonstrations still show shortcomings in terms of fine dexterity compared to Figure 02.
# Comparatif des architectures d'actionneurs
# Figure 02 vs Tesla Optimus Gen 2
class Actionneur:
def __init__(self, nom, couple, puissance):
self.nom = nom
self.couple = couple # en Nm
self.puissance = puissance # en Watts
# Approche Figure AI : Haute performance, custom
figure_actuator = Actionneur(
nom="Figure Custom Rotary",
couple=150, # Couple élevé pour des mouvements fluides
puissance=500
)
# Approche Tesla : Optimisation des coûts, design maison
tesla_actuator = Actionneur(
nom="Tesla Gen 2 Rotary",
couple=110, # Couple inférieur mais suffisant pour des tâches basiques
puissance=300
)
print(f"Avantage Figure : {figure_actuator.couple - tesla_actuator.couple} Nm de couple en plus.")
print(f"Avantage Tesla : Coût de production estimé 40% inférieur grâce à l'économie d'échelle.")
Optimus Strengths and Weaknesses
According to The Verge reports, Tesla's major asset is its ability to collect data. Tesla's autonomous car network indirectly feeds Optimus's training in terms of spatial perception. Nevertheless, manipulating a steering wheel is not the same as manipulating a fragile shipping box. Optimus's current weakness lies in its hands, which seem less capable of fine, delicate manipulation than those of Figure 02.
Unitree G1: The Chinese Prodigy of Democratization
If America dominates AI models, China strikes hard on hardware. Unitree, known for its quadruped robots (robot dogs), surprised everyone by unveiling the G1, a humanoid sold at a base price of PH_prix_unitree_g1_PH dollars.
A Disruptive Price for the Developer Ecosystem
The Unitree G1 does not have the finesse of Figure 02, but it has a different goal: to become the "Android of robotics." By selling it so cheaply, Unitree targets research labs, universities, and independent developers to create an application ecosystem.
{
"unitree_g1_specs": {
"prix_public": "16 000 $",
"hauteur": "1.27 m",
"poids": "35 kg",
"degrés_de_liberte": "23 articulations articulées",
"actionneurs": "Moteurs articulés à entraînement direct (max 8.5 Nm)",
"perception": "Caméras Intel RealSense D435 + LiDAR optionnel",
"compute": "Nvidia Jetson Orin NX",
"specialite": "Sim-to-Real Transfer, marche adaptative, arts martiaux simulés"
}
}
The Sim-to-Real Bet
Unitree excels in skill transfer. Thanks to simulation environments like Isaac Sim (Nvidia), the G1 learns to walk, fall and get back up, or even dodge attacks (as shown in their Kung-Fu videos), before transcribing this data into its physical body. This is a highly effective approach for locomotion, but one that still requires improvements for complex tool manipulation.
1X Neo: The Soft Alternative for the Domestic Market
While the others fight for factories, the Norwegian startup 1X Technologies (strongly backed by OpenAI) is betting on a different niche with the 1X Neo: the home and care market.
Design and Home Integration
The 1X Neo abandons the industrial aesthetic for a soft design, with an expressive face and a structure that inspires trust. It is designed to move silently, carry fragile objects, and interact with vulnerable humans (the elderly, children).
- Maximum speed: PH_vitesse_max_neo_PH km/h (designed for indoor safety, compared to 3 to 5 km/h for the others)
- Weight: 30 kg (the lightest of the four)
- Autonomy: 2 to 4 hours depending on household tasks
- Estimated price: Premium positioning, estimated between $25,000 and $40,000 for the initial phases.
The strength of 1X Neo lies in its software. As one of the first robots to natively integrate OpenAI's AI models publicly, it understands complex linguistic requests extremely well ("Put the red objects on the top shelf, but be careful with the vase").
Head-to-Head Comparison: Prices, Dates, and Investors
To see things more clearly, let's examine these four players through the cold lens of business data and the market timeline.
-- TABLEAU COMPARATIF : LA COURSE AUX HUMANOÏDES 2024-2025
-- Source : Synthèse de données The Verge, communiqués startups
CREATE TABLE race_humanoide (
robot_nom VARCHAR(50),
entreprise VARCHAR(50),
prix_estime_usd INT,
date_commercialisation_prevue VARCHAR(100),
investisseurs_principaux TEXT,
avantage_competitif TEXT
);
INSERT INTO race_humanoide VALUES
(
'Figure 02',
'Figure AI',
[[prix_cible_figure]],
'Déploiement en usine BMW fin 2024 / Commercialisation grand public 2026-2027',
'Nvidia, OpenAI, Microsoft, Jeff Bezos',
'Dextérité fine et intégration VLM poussée'
),
(
'Optimus Gen 2',
'Tesla',
20000,
'Usage interne Tesla fin 2024 / Vente externe 2025-2026',
'Capitalisation boursière Tesla (Fonds propres)',
'Écosystème de fabrication et coûts réduits à l''extrême'
),
(
'Unitree G1',
'Unitree Robotics',
[[prix_unitree_g1]],
'Déjà en pré-commande pour développeurs (Livraisons 2024)',
'Fonds chinois locaux, ventes directes B2C/B2B',
'Prix de rupture et écosystème open-source/developers'
),
(
'1X Neo',
'1X Technologies',
30000,
'Tests en garderie/nursing 2024 / Commercialisation 2026',
'OpenAI Startup Fund, EQT Ventures',
'Design sécurisant et spécialisation domestic-care'
);
-- Requête pour visualiser le rapport qualité/prix
SELECT
robot_nom,
prix_estime_usd,
avantage_competitif
FROM race_humanoide
ORDER BY prix_estime_usd ASC;
Technical Analysis: Who Wins the AI Marathon?
Having a sleek mechanical body is useless without a suitable brain. The real battle is taking place at the level of the embedded AI architecture. Let's take the example of a simple task for a human: sorting recyclable waste.
# Architecture type de résolution de problème chez un humanoïde moderne
# Comparaison de la latence et de la gestion des erreurs
import time
class RobotBrain:
def executer_tache(self, tache):
start_time = time.time()
# 1. Perception (Vision + Audio)
perception_data = self.perception.process_environment()
# 2. Raisonnement de haut niveau (LLM/VLM)
plan_action = self.llm.generate_plan(
tache=tache,
context=perception_data
)
# 3. Planification motrice (Low-level control)
trajectoire = self.motor_planner.calculate(
plan=plan_action,
robot_state=self.kinematics.get_state()
)
# 4. Exécution et retour d'erreur (Feedback loop)
resultat = self.actuators.execute(trajectoire)
if resultat.error:
# Le robot doit réajuster en temps réel
self.llm.replanify(resultat.error)
end_time = time.time()
return end_time - start_time
# Scénario : Figure 02 vs Unitree G1 sur une tâche imprévue
figure_brain = RobotBrain(model="Figure-OpenAI-VLM", compute="3x Orin")
unitree_brain = RobotBrain(model="Unitree-Loco-LLM", compute="1x Orin NX")
# Figure 02 aura une latence plus faible sur le raisonnement visuel grâce à ses 3 Orin
# Unitree G1 excellera sur l'adaptation posturale (sim-to-real) mais sera plus lent
# dans la compréhension fine de l'objet (ex: est-ce que ce plastique est trop sale pour être recyclé ?)
This is where Figure 02 takes a decisive theoretical advantage thanks to its close partnership with OpenAI. Their engineers managed to integrate neural networks directly into the low-level control loops, allowing the robot to react to an object slipping from its fingers in a few milliseconds, without having to make a round trip to a Cloud server.
Tesla, on the other hand, is betting on its pure vision technology (cameras only, no LiDAR) derived from its cars. While this reduces costs, it makes the Optimus robot potentially vulnerable in very low-light environments or when facing transparent objects, a challenge that Figure 02 circumvents through a combination of sensors.
Current Limitations: The Reality Behind Viral Videos
It's easy to get carried away by the impressive video edits on X (formerly Twitter) or YouTube. However, The Verge regularly points out that we need to separate laboratory demonstrations from commercial viability.
The Problem of Endurance and Maintenance
None of these robots are currently capable of working 8 hours straight. The 5-hour battery life announced by Figure 02 is an ideal average. In a real factory environment, repeatedly lifting 20 kg drains the battery much faster. Furthermore, high-torque actuators (especially those from Figure and Optimus) generate a lot of heat and require active cooling systems or regular breaks.
The Dexterity Glass Ceiling
We are excellent at making robots walk. Dynamic locomotion (jumping, running, recovering from a push) is largely solved thanks to reinforcement learning. But fine manipulation remains the Holy Grail. Buttoning a shirt, handling flexible electronic cables, or separating spaghetti on a plate are tasks that require tactile perception and precision that neither Figure 02 nor Optimus reliably master 100% today.
The Supply Chain Puzzle
Building 10,000 humanoid robots is not like manufacturing 10,000 smartphones. Custom actuators, harmonic reducers (like those from Harmonic Drive), and custom force sensors have very long production lead times. Tesla is best positioned to solve this problem thanks to its foundry experience (Giga Press), but other players could hit a production wall by 2026.
Summary of Strengths and Weaknesses
- Figure 02: Leader in dexterity and AI integration (OpenAI). Best for complex logistical tasks. Weakness: High production cost and reliance on cutting-edge external components.
- Tesla Optimus: King of economies of scale and mass production. Weakness: Hand dexterity lagging behind competitors, reliance on purely camera-based vision.
- Unitree G1: The champion of price and locomotion. Ideal for research and rapid prototyping. Weakness: Less sophisticated AI for complex task reasoning, limited payload.
- 1X Neo: The specialist in domestic and care environments. Most socially acceptable design. Weakness: Raw physical performance far below competitors, slow speed.
Conclusion: Who Really Wins?
If we look at the race through the prism of immediate technological maturity and software integration, Figure 02 wins the 2024 round. Their robot is currently the most advanced at performing useful, non-pre-scripted work in a real environment, as proven by their current tests at BMW.
However, if we look at the race in the long term (2028-2030), Tesla Optimus is the structural favorite. Tesla's ability to design its own chips, cast its own parts, and produce at a rate of one million units per year is an asset no one else possesses. If Tesla manages to catch up on its software lag in manipulation, Optimus will become the industry standard.
Finally, let's not forget Unitree G1, which could well become the "Linux of robotics": present in labs everywhere, training the next generation of engineers, and infiltrating from the ground up into niches that the big players neglect.
The humanoid robot race is just beginning. The coming months, with the first real deliveries outside of laboratories, will tell if these promises hold up. Stay tuned, watch the weight updates of these models, and get ready: the era of human-machine cohabitation is accelerating.
Ready to dive deeper into AI applied to robotics?
Discover our practical guides on AI-master.dev to learn how to train your own computer vision models and understand the architectures that will shape the world of tomorrow. [Join the AI-master.dev newsletter so you don't miss any of the latest updates from Skynet Watch].