Chapter 1
The State of the Field in One Page
3 sections · about 5 minutes
Robotics in 2026 is in the middle of a bet that has not yet paid off and has not yet failed.
The bet is that the recipe that worked for language — a single large model, pretrained on a very large and diverse corpus, then adapted cheaply to specific tasks — will work for physical action. The models are called VLAs (vision-language-action models). The corpus is the problem: there is no internet of robot behaviour, and there is no obvious way to make one.
Everything else follows from that. The academic labs have split into those trying to manufacture the corpus (Stanford's IRIS and REAL, MIT's Improbable AI, Toyota Research Institute), those arguing the corpus is the wrong idea and structure should replace scale (MIT's LIS group, the RAI Institute, most of the control-theory establishment), and those building the sensors and mechanisms everyone else will eventually need (GelSight, BDML, CHARM, the Movement Lab). The companies have split into those selling bodies (Figure, Unitree, Agility, UBTech), those selling brains (Physical Intelligence, Skild, Generalist, Google DeepMind), those selling the picks and shovels (NVIDIA, Hugging Face, the data-labelling industry), and those quietly selling actual working robots into actual warehouses (Dexterity, Ambi, Path, Chef, RightHand).
Four numbers frame the whole field.
- ~3,000 hours. The size of the largest purpose-built real-robot manipulation dataset in existence, AgiBot World — 1,001,552 trajectories, 2,976.4 hours, collected by 100+ robots in a purpose-built 4,000 m² facility.
- ~100,000 years. Ken Goldberg's estimate, in Science Robotics, August 2025, of the human-equivalent time represented by the text and image corpora used to train frontier VLMs. The ratio between these first two numbers is the entire problem.
- 1,250 hours. The runtime of the most-cited real humanoid deployment in the West — Figure 02 at BMW Spartanburg — over eleven months, loading 90,000+ parts at one station. That is roughly seven months of one human's working time.
- 73.6%. The share of Unitree's humanoid revenue that came from research and education buyers in the first nine months of 2025, per its IPO prospectus. The only company with audited humanoid revenue is mostly selling to people who study robots, not people who use them.
Hold those four numbers in mind and most of the rest of this book is commentary.
1.1 A compressed timeline, 2022–2026
| Date | Event | Why it mattered |
|---|---|---|
| Dec 2022 | Google RT-1 | First serious "transformer for robot actions" at scale |
| Mar 2023 | RSS: Diffusion Policy (Chi, Song, Tedrake et al.) | Generative modelling of action sequences becomes the default architecture |
| Jul 2023 | RT-2 | A VLM fine-tuned to emit actions — the VLA template |
| Jul 2023 | ALOHA (Zhao, Finn) | A <$20k bimanual teleop rig; democratises data collection |
| Oct 2023 | Open X-Embodiment | 21 institutions pool 1M+ trajectories, 22 embodiments; positive cross-embodiment transfer demonstrated |
| Jan 2024 | Mobile ALOHA | Adds a mobile base; the viral moment for low-cost imitation |
| Feb 2024 | UMI (Chi, Song et al.) | Handheld gripper decouples data collection from robot ownership |
| Aug 2024 | Amazon hires Covariant's founders | The first "reverse acquihire" in robot foundation models |
| Oct 2024 | Physical Intelligence π0 | The first credible "one policy, many robots" commercial model |
| Nov 2024 | Stanford Robotics Center opens | Academic robotics gets a shared physical plant |
| Mar 2025 | NVIDIA Isaac GR00T N1 | An open humanoid foundation model; NVIDIA positions as the platform |
| Mar 2025 | Gemini Robotics | DeepMind folds robotics into its frontier model line |
| Apr 2025 | π0.5 | Open-world generalisation to unseen homes; diminishing returns above ~100 homes |
| Jul 2025 | TRI's Large Behavior Models paper | The field's first rigorous, blind, statistically-controlled evaluation — and a sobering one |
| Aug 2025 | Goldberg's "100,000 Year Data Gap" | The scaling-skeptic case gets its canonical citation |
| Sep 2025 | Figure Series C, $39B | Peak valuation for a pre-revenue humanoid company |
| Sep 2025 | Brooks: "Why Today's Humanoids Won't Learn Dexterity" | The tactile-first critique gets its canonical citation |
| Nov 2025 | π*0.6 + RECAP; Generalist GEN-0 | RL-from-experience, and the first published embodied scaling-law claim |
| Jan 2026 | Skild AI, $1.4B at $14B+ | The "omni-bodied" thesis gets funded at scale |
| Feb 2026 | Intrinsic absorbed into Google | Alphabet consolidates physical AI |
| Mar 2026 | Unitree STAR Market IPO filing | First audited humanoid P&L |
| Apr 2026 | π0.7; Generalist GEN-1 | Compositional generalisation claims; 500k hours of human wearable pretraining with zero robot data |
| Jun 2026 | Agility SPAC ($2.5B); Neura Series C (up to $1.4B); Sanctuary pivots hardware-agnostic | Late-cycle financing signals |
| Jul 2026 | Walden Robotics launches (Tedrake, $1.1B); Gemini Robotics 2; FCC adds mobile robots >2kg to the Covered List | Geopolitics arrives |
1.2 The stack, from silicon to task
Understanding any lab or company in this book means locating it on this stack. Almost nobody spans more than three layers well.
- Actuators and transmissions — 40–60% of humanoid bill-of-materials cost per McKinsey. Quasi-direct-drive, harmonic reducers, tendon drives. Sangbae Kim's lab, Steve Collins' lab, LG Axium, Hyundai Mobis, Unitree.
- Hands and end effectors — parallel jaws, suction, multi-fingered, tendon-driven. 1X (25-DOF), Apptronik SharpaWave (22-DOF), Sanctuary hydraulics, RightHand hybrids.
- Sensing — cameras, depth, proprioception, and the missing modality, touch. GelSight/Adelson, Kennedy's DenseTact, Meta Digit 360, Amazon Vulcan.
- Low-level control — whole-body control, MPC, operational-space formulation, balance at 1 kHz. Khatib, Boston Dynamics, Biomimetic Robotics, Figure's S0.
- Policy — the learned visuomotor layer. Diffusion policies, flow matching, action chunking. Tedrake/TRI, Song, Finn, Physical Intelligence.
- Reasoning and task planning — VLMs, embodied reasoning, TAMP. Gemini Robotics ER, Figure's S2, MIT LIS, Jiajun Wu.
- Data and evaluation — collection interfaces, datasets, simulators, benchmarks. Everyone claims this layer; almost nobody has solved it.
1.3 Glossary of terms used throughout
- VLA — Vision-Language-Action model. A network mapping images plus a language instruction to robot actions.
- Action chunking — predicting a short sequence of future actions rather than one step, which reduces compounding error. Introduced at scale by ACT (ALOHA).
- Diffusion policy — modelling the action distribution with a denoising diffusion or flow-matching process; now the default for multimodal manipulation behaviour.
- Cross-embodiment — training one model on data from multiple different robot bodies.
- Embodiment gap — the mismatch between the body that generated data (a human, another robot) and the body executing it.
- Sim-to-real gap — the divergence between simulated and real physics, worst for contact, friction, and deformables.
- Teleoperation — a human directly driving the robot, usually the source of demonstration data.
- Perioperation — Improbable AI's term for collecting data with a passive exoskeleton mechanically coupled to the robot's hand, so a human's demonstration is already in the robot's action space.
- Data flywheel — the claim that deployed robots generate the data that improves the next model, which improves deployment.
- BOM — bill of materials; the hardware cost of a robot.
- RaaS — robots-as-a-service; renting the robot per month rather than selling it.
- MTBF — mean time between failures. The number almost nobody in humanoid robotics reports.
- Whole-body control / loco-manipulation — coordinating locomotion and manipulation as one problem rather than two.