Detailed content
🚶 Why humanoids: a world built for humans
The humanoid thesis is simple and powerful: the world has already been built for the human form. Door handles, stairs, tools, workbenches, and vehicles assume two hands, two feet, and person height. A humanoid robot operates within this worldwide stock of infrastructure— brownfield — without renovating it, unlike fixed arms that require dedicated cells.
💡 The underlying bet
Generality of form (humanoid) × generality of brain (foundation model VLA) = a single product addressing countless tasks. This is the convergence driving tens of billions in capital into the sector since 2024.
📊 Why the human form is difficult—and worth it
- ~30-50 DOF — a humanoid has an order of magnitude more degrees of freedom than an arm, requiring whole-body control.
- Dynamic balance — A biped is inherently unstable; manipulating while walking is the real test of usefulness.
- Enormous TAM — the economic argument is to address generic physical work, not just one station.
Generality of form
1 body, N environments.
Brownfield
Operates on existing infrastructure.
Dexterity
Dexterous multi-finger hands.
DOF
High degrees of freedom.
🟢 NVIDIA GR00T (N1/N1.5): dual foundation model
GR00T is NVIDIA’s open foundation model for humanoids. The N1 ushered in a cross-embodiment foundation model with dual system — a slow reasoning module (System 2, VLM-based) and a fast action module (System 1, diffusion-based). The N1.5 refined instruction following and generalization. NVIDIA couples this with the Isaac GR00T + Cosmos pipeline to generate synthetic data and neural-trajectories at scale.
GR00T N1 — pilha de dados (pirâmide)
topo: poucas demos reais (teleoperação) ← fidelidade
meio: dados sintéticos (Isaac/Cosmos) ← escala física
base: vídeo humano web + neural trajectories ← semântica / amplitude
=> cross-embodiment: 1 modelo treinado para muitos corpos⚡ Why it matters strategically
NVIDIA doesn’t manufacture humanoids — it provides the brain + data factory + silicon. Several OEMs (1X, Agility, Boston Dynamics, Unitree, and others) adopt GR00T/Isaac, positioning NVIDIA as the ecosystem’s horizontal layer, analogous to what CUDA was for deep learning.
GR00T N1/N1.5
Open foundation model for humanoids.
Dual system
Reasoning + action.
Cosmos / Isaac
Large-scale synthetic data.
Cross-embodiment
1 model, many bodies.
🔵 Figure Helix: hierarchical VLA (System 1/2)
Helix, from Figure, is a hierarchical VLA that ran onboard a humanoid. The System 2 (internet VLM, ~7-9 Hz) understands the scene and instruction and emits a latent vector; o System 1 (visuomotor policy, ~200 Hz) decodes this latent into continuous control of the torso, head, and dexterous hands. It was one of the first to control a humanoid's entire upper body with a single VLA—and to run two robots collaborating with the same model.
System 2 — think (slow)
A generalist VLM interprets "put away the dishes"; it produces a latent intention without needing to run at 200 Hz.
System 1 — act (fast)
Visuomotor transformer consumes the latent and generates high-frequency commands for hands and torso.
💡 The conflict that dual-system architecture resolves
A large VLM is smart but slow; a fast controller is agile but dumb. Separate the two and connect them through a latent vector offers the best of both: generalist reasoning + a high-frequency control loop. It's the convergent architecture of leading humanoids.
Helix
Figure’s hierarchical VLA.
System 1/2
Fast + slow coupled.
Latent vector
Bridge between the systems.
Whole-upper-body
Torso + hands with 1 VLA.
🔴 Tesla Optimus: integration with the vision/FSD stack
Optimus is Tesla's bet on a humanoid reusing FSD's AI infrastructure: end-to-end vision networks, proprietary compute, and a culture of large-scale data collection. The strategy is to start with the in-house manufacturing (moving parts in their own factories), generating vertically integrated teleoperation data before scaling outward.
✓ Tesla’s advantages
- ✓Reuse of the stack of end-to-end vision and FSD data.
- ✓Scale of proprietary manufacturing and actuators designed in-house.
- ✓Captive internal demand to close the data loop.
✗ Open challenges
- ✗Manipulation dexterous robust is still the bottleneck (hands).
- ✗Historically optimistic announced timelines.
- ✗Generalization beyond the factory remains unproven at scale.
Optimus
Tesla humanoid.
FSD reuse
Shared end-to-end vision.
Industrial scale
The factory as the first customer.
Custom actuators
Vertical hardware integration.
🟠 Unitree and open hardware: G1/H1
A Unitree changed the research game by launching low-cost humanoids: the H1 (full-size, agile locomotion) and the G1 (~1.27 m, starting at ~$16k). Suddenly, labs without million-dollar budgets could run policies on real hardware. The result was an explosion of papers on locomotion and loco-manipulation using G1/H1 as the standard platform.
📊 Why cost matters
- ~10-50× cheaper than comparable Western research platforms.
- Open SDK and ROS2 — a low barrier to integrating VLA policies and locomotion RL.
- Standardization — Comparability across papers accelerates the entire field.
⚡ Strategic tip
Accessible hardware shifts the bottleneck in the field from "having a robot" to "having data and algorithms." Those who master data collection and architecture—not those who manufacture the arm—capture more value. That’s how open hardware accelerated the frontier.
Unitree G1/H1
Low-cost humanoids.
~US$ 16k
G1 base price.
Open SDK
Easy integration.
Default platform
Basis for many papers.
⚖️ Locomotion + manipulation: whole-body control
The humanoid frontier isn't walking or grasping in isolation—it's do both at the same time. O whole-body control coordinates gait, balance, and arms in a single controller, usually with RL for robust locomotion across varied terrain and a coupled manipulation policy. Manipulating while balancing is what separates a viral demo from real economic utility.
Locomotion RL in simulation
Robust locomotion policy trained on thousands of randomized terrains (M3.1 sim-to-real).
Balance under perturbation
Control the center of mass while the arms apply external force (pushing, carrying).
Coupled loco-manipulation
Walk to a cabinet, open it, and put an object away: skills chained together in a single body.
Whole-body control
Coordinate the whole body.
Locomotion RL
Robust locomotion.
Balance
Center of mass under force.
Loco-manipulation
Walk + manipulate together.
✅ Module summary
Next module
3.3 — Hybrid Architectures: LLM Planner + VLA Executor. How the dual system generalizes to long-horizon planning.