Chaos and Order

💡 왼쪽 원문을 읽으면서 오른쪽에 따라 써보세요. Tab 키로 힌트를 받을 수 있습니다.

Introduction — Why Robots, Why Now
Part 1 — The Company Map: Hardware × Brain × Market
Part 2 — The VLA Model Lineage: From RT-2 to GR00T N2
Part 3 — The Software Stack: What Engineers Actually Touch
Part 4 — The Engineer's Way In
Closing Thoughts
References

Introduction — Why Robots, Why Now

Let me start with a single number: the forecast for global humanoid shipments in 2026 is 50,000+ units — roughly 7x year-over-year. More than 140 companies are building humanoids priced from $16,000 to $250,000, and this is exactly where the capital and talent of the LLM boom are rushing in search of "the next substrate." The trigger for the shift was not hardware but the brain — with the arrival of VLA (Vision-Language-Action) models that carry LLM/VLM capabilities into robot actions, the decades-old limitation that "robots only do what they are programmed to do" has started to crumble.

This post draws the map in three layers — companies (who is building), models (how the brain has evolved), and the stack (what engineers actually work with).

Part 1 — The Company Map: Hardware × Brain × Market

United States — The Battleground of Capital and Foundation Models

Figure — The capital leader of this arena. A large round led by OpenAI put its valuation at roughly $39 billion. The heart of the strategy is not hardware but Helix, its in-house VLA model — a bet that "the body is the vehicle; the moat is embodied AI." Figure 03 is the current platform.
Tesla — Optimus bets on manufacturing scale rather than outside funding. Gen 3 targets mass production in the summer of 2026 — the proving ground for the scenario where "the company that owns car factories stamps out robots."
Physical Intelligence (π) — It does not sell robots. It is a software company building a general-purpose brain (π0, π0.5) that can be transplanted into any robot, and its diffusion-based action generation leads the dexterity benchmarks. The contrast with OpenVLA: a closed model, accessible only through partnerships.
1X — With its home humanoid NEO, it takes on "inside the home" — the hardest environment of all — head-on. Apptronik (Apollo) and Agility Robotics (Digit) are stacking up commercial deployments in logistics and warehousing, the most realistic market. Boston Dynamics shed hydraulics with the electric Atlas and is converting from research royalty into a commercial player.

China — Price and Volume

Unitree — The price breaker of this arena. Leveraging the mass-production muscle it built with quadrupeds, it pushes the G1 humanoid out at a low price point and shipped more than 5,500 units in 2025 alone — the world leader by volume. At a valuation of roughly $1.3 billion its capital is one-thirtieth of Figure's, but the strategy — "ship cheap, ship many, harvest the data" — is formidable. Behind it, UBTech, Fourier, AgiBot, and others are chasing hard on the back of China's manufacturing ecosystem.

Korea — A Manufacturing Powerhouse Quietly Joins the Fight

Rainbow Robotics — Heir to the KAIST HUBO lineage; with Samsung Electronics becoming its largest shareholder, it is now the pivot of the group's robotics strategy. Doosan Robotics, a collaborative-robot (cobot) powerhouse, keeps taking over industrial floors, and counting Boston Dynamics under Hyundai Motor Group, "the robot portfolio held by Korean capital" is larger than you might think.

The thing to watch is how three axes combine — hardware (actuators, hands), brain (VLA), and market (factory → logistics → home). Figure bets on vertically integrating the brain, Tesla on manufacturing, Unitree on price, and π on selling the brain horizontally — each staking a different axis.

Part 2 — The VLA Model Lineage: From RT-2 to GR00T N2

A VLA folds "look at the image (Vision), understand the instruction (Language), output motor commands (Action)" into a single model. Trace the lineage and today's position comes into focus.

RT-2 (Google DeepMind, 2023)    Proof of concept wiring a VLM into robot actions — "internet knowledge reaches the arm"
OpenVLA (2024)                  7B open-source VLA — a starting point anyone can download and fine-tune
π0 (Physical Intelligence)      Diffusion-based continuous actions — the new bar for dexterity, closed
Helix (Figure)                  "Fast reflexes + slow reasoning" dual system, humanoid upper-body control
GR00T N1 → N1.6 → N2 (NVIDIA)   Open-weight humanoid foundation models — N1.6 is 3B parameters,
                                N2 the first large VLA to make 30+ DoF humanoids its primary design target

Only three technical currents need remembering. First, the evolution of action representations — from discrete tokens (RT-2) to diffusion/flow matching (π0), toward smooth and precise continuous control. Second, data strategy — real-robot data is expensive, so a two-stage recipe has become standard: pre-learn the world's dynamics from web video and egocentric footage (V-JEPA 2, GR00T's latent-action pretraining), then align on robot data. Third, fusion with world models — world foundation models like NVIDIA Cosmos are being absorbed into VLA backbones, evolving toward "imagine the consequences of an action, then move." This data-pipeline instinct follows exactly the same grammar as the LLM data preprocessing post — only the raw material has changed, from text to video and trajectories.

Part 3 — The Software Stack: What Engineers Actually Touch

Layer               Tool                       Role
──────────────────  ─────────────────────────  ────────────────────────────
Simulation          NVIDIA Isaac Sim / Lab     Photoreal physics sim + RL framework
                    MuJoCo                     Research-standard lightweight physics engine (DeepMind)
Middleware          ROS 2                      The Linux of robotics — node/topic/action communication
Foundation model    Isaac GR00T (N series)     Open-weight humanoid VLA
Training framework  LeRobot (HuggingFace)      The Transformers of robotics — datasets, policies, real-robot training
Data                Open X-Embodiment          A joint dataset of trajectories from dozens of robots

The thing worth flagging is the standardization of the sim-to-real pipeline — build a policy with massively parallel reinforcement learning in Isaac Lab, close the reality gap with domain randomization, then transplant it onto real hardware; that flow has become the de facto textbook. And the place where all this training runs is precisely the multi-GPU cluster — robot AI is the next-door neighbor of LLM infrastructure.

Part 4 — The Engineer's Way In

"Don't robotics companies only hire robotics PhDs?" — Not anymore. A robotics company in the VLA era is, in practice, a giant ML infrastructure + data engineering organization, and half of any job posting reads like the skills in the rising-roles knowledge map. A realistic three-step path:

Start with LeRobot — train a policy on public datasets from your laptop, and if budget allows, go all the way to real hardware with a low-cost robot arm (the SO-ARM family). It is the robotics edition of "do it end-to-end, once."
Read three papers closely — OpenVLA (the open-source reference point), π0 (diffusion actions), and GR00T N1 (humanoid foundation). Those three pin down the coordinate system of the current terrain.
Simulation skills — run a reinforcement learning pipeline once with Isaac Lab or MuJoCo, and you will hold your own in a robotics team interview.

To a working ML engineer, robotics is less "a new field" than one more deployment environment with strange distributions and brutal latency requirements — and that point of view is exactly the one this industry is thirsting for right now.

Closing Thoughts

The robotics industry of 2026 is crossing from the era of "hardware demos" into the era of "brains and volume." Figure's $39 billion and Unitree's 5,500 units are two strategies on the same board, and the board's lingua franca is VLA. If you have been working on LLM infrastructure — your skills are already half robotics skills. The physical world will teach you the other half.