Our last deep dive covered how text, image, audio, and video are consolidating into single any-to-any foundation models. This post covers what happens when that consolidation reaches past the screen — into robots that see a countertop, reason about the dishes on it, and actually pick one up without dropping it. That's physical intelligence: the discipline of building AI that perceives, plans, and acts in real, physical environments. It's also the name Silicon Valley's most-funded robotics startup chose for itself in 2024 — a deliberate statement that "action" is simply the next modality to fold into the foundation-model paradigm. This article covers the field end-to-end: the history, the architectures, the companies racing to build generalist robots, the money pouring in, and the open research problems that separate today's viral demo videos from a robot that reliably does your laundry.
🕰️ From Hand-Coded Control to Generalist Robots
Robotics is one of the oldest branches of AI — and one of the slowest to benefit from the deep learning revolution, because physical action doesn't come with an internet's worth of labeled training data the way text and images did. Here's how the field got from programmable arms to models that generalize across tasks, robots, and environments:
Classical / Pre-Learning Era
Generalist VLA / Foundation Model
⚙️ Era 1 — Classical Robotics (1961–2015)
1961
Classical Era
1961
Unimate
General Motors · The first industrial robot arm
ApproachHydraulically actuated arm executing fixed, pre-programmed sequences on a GM die-casting line — zero perception, zero learning, pure repeatable mechanics.
LimitationCould only perform the exact motion it was programmed for. Any change to the environment or task required manual reprogramming by an engineer.
Classical EraFoundational
1986
Architecture
1986
Behavior-Based Robotics (Subsumption Architecture)
Rodney Brooks, MIT · Reactive layered control
ApproachReplaces centralized world-modeling with layered reactive behaviors that respond directly to sensor input — "the world is its own best model."
ImpactInfluenced a generation of mobile robots (including early Roomba) that prioritized robustness over precise planning — a philosophical precursor to today's end-to-end learned policies.
ArchitectureMIT
2005
Classical Era
2005
Boston Dynamics BigDog
Boston Dynamics · Dynamic legged locomotion via model-based control
ApproachHand-engineered dynamics models and control loops (not learned) achieve startlingly robust quadruped balance and terrain traversal — a triumph of classical control theory, not machine learning.
LimitationEvery new behavior required painstaking manual controller engineering — no generalization to novel tasks without new hand-built control code.
Classical EraLocomotion
🧠 Era 2 — Deep Reinforcement Learning for Robotics (2015–2021)
2015
Learned Policy
2015
End-to-End Visuomotor Policies
Sergey Levine, Chelsea Finn et al. (UC Berkeley) · Pixels-to-torques deep learning
ApproachTrains a single CNN policy to map raw camera pixels directly to joint torques for manipulation tasks, bypassing hand-engineered perception and control pipelines entirely.
ImpactFirst credible demonstration that end-to-end learning could replace decades of modular robotics engineering — the conceptual seed of everything that followed.
Learned PolicyBerkeleyFoundational
2017
Architecture
2017
Domain Randomization & Sim-to-Real
OpenAI, UC Berkeley · Training in simulation, deploying in reality
ApproachRandomizing simulator textures, lighting, physics parameters during training forces policies to learn features robust enough to transfer to the real world without ever training on real data.
ImpactMade simulation a viable, scalable data source for robotics — critical given how expensive and slow real-world robot data collection is.
ArchitectureSimulation
2019
Learned Policy
Jul 2019
OpenAI Rubik's Cube Hand
OpenAI · Dexterous in-hand manipulation via large-scale RL + domain randomization
ApproachA robotic hand trained almost entirely in simulation learns to manipulate a Rubik's Cube one-handed, transferring to physical hardware via aggressive domain randomization.
ImpactWidely cited proof that scale (compute + randomized simulation) could substitute for hand-engineered dexterity — foreshadowed the "scale is all you need" thesis later applied to VLA models.
Learned PolicyOpenAIDexterity
🌉 Era 3 — Foundation Models Meet Robotics (2022–2023)
2022
Architecture
Apr 2022
SayCan
Google Research · Grounding LLM plans in robot affordances
ApproachUses a large language model to propose candidate high-level actions from natural language instructions, then scores each by a learned "affordance" value function reflecting what the robot can actually do right now.
ImpactFirst credible bridge between LLM reasoning and physical robot execution — the direct conceptual ancestor of vision-language-action models.
ArchitectureGoogleLLM-Grounding
2023
Generalist VLA
Dec 2022 – Jul 2023
RT-1 → RT-2 ⭐
Google DeepMind · First Vision-Language-Action foundation models
ApproachRT-2 fine-tunes a web-scale vision-language model (based on PaLI-X/PaLM-E) to output robot actions as text tokens — treating control as just another "language" the model learns to speak.
ImpactDemonstrated that internet-pretrained VLM knowledge (semantic understanding of objects, concepts never seen in robot data) transfers directly into physical manipulation — coined the term "Vision-Language-Action" model and defined the architecture every major lab now follows.
Generalist VLAGoogleMilestone
Open Source
Oct 2023
Open X-Embodiment / RT-X
Google DeepMind + 34 academic labs · Cross-embodiment robot dataset & model
ApproachPools 1M+ real robot trajectories across 22 different robot embodiments from 34 labs worldwide into one unified dataset and training run.
ImpactProved that training jointly across many different robot bodies improves performance on each individual robot — the "cross-embodiment transfer" finding that underlies every generalist robot model since.
Open SourceDatasetCross-Embodiment
🚀 Era 4 — The Generalist Robot Race (2024)
2024
Company Founding
Nov 2024
Physical Intelligence Founded ⭐
Physical Intelligence (Ï€) · Ex-Google DeepMind, Stanford, UC Berkeley robotics researchers
ApproachFounded by Karol Hausman, Sergey Levine, Chelsea Finn, Brian Ichter and others with the explicit thesis that a single generalist foundation model — not per-task engineering — will produce robots that work in unstructured, real-world environments.
ImpactRaised the largest seed-adjacent robotics round in history ($400M Series A at a $2.4B valuation, backed by OpenAI and Jeff Bezos) — a strong signal that frontier AI capital believes physical intelligence is the next consolidation frontier.
Company FoundingMilestone
Generalist VLA
Nov 2024
π0 (pi-zero) ⭐
Physical Intelligence · Flow-matching VLA generalist policy
ApproachCombines a pretrained vision-language backbone with a flow-matching action-generation head, trained across 7 different robot platforms performing laundry folding, table bussing, box assembly, and dozens of other dexterous tasks.
ImpactThe clearest public demonstration yet of a single model performing genuinely dexterous, long-horizon household tasks across multiple robot embodiments — set the architecture reference point (VLA + flow-matching action head) the rest of the field is now converging toward.
Generalist VLAMilestoneFlow Matching
Generalist VLA
2024
Figure Helix
Figure AI · Full-body VLA control for humanoid robots
ApproachA dual-system VLA architecture — a slow "System 2" reasoning model plans, a fast "System 1" reactive model executes — running entirely onboard the Figure 02 humanoid to control hands, arms, and full-body motion in real time.
ImpactFirst humanoid VLA to demonstrate two robots collaborating on a shared task (unloading groceries) using natural language coordination — a preview of multi-robot generalist deployment.
Generalist VLAHumanoid
Open Source
2024
NVIDIA GR00T / Isaac Lab
NVIDIA · Open humanoid foundation model + simulation infrastructure
ApproachGR00T N1 provides an open foundation model for humanoid control, paired with Isaac Lab/Sim simulation infrastructure and synthetic data generation pipelines that any robotics team can build on.
ImpactPositioned NVIDIA as the "compute + tooling" layer beneath nearly every humanoid startup — the CUDA of embodied AI, capturing value regardless of which robot company wins.
Open SourceNVIDIAInfrastructure
🌍 Era 5 — Commercialization & Scale-Up (2025–2026)
2025
Generalist VLA
2025
Gemini Robotics ⭐
Google DeepMind · Gemini-native embodied reasoning and control
ApproachExtends the natively multimodal Gemini backbone with a robotics action-output head — the same model that reasons over video and audio now also reasons over physical affordances and outputs low-level control.
ImpactThe clearest statement yet that "action" is being folded into the same consolidated foundation-model stack as every other modality, rather than remaining a separate specialist system — directly extending the "Great Consolidation" thesis into the physical world.
Generalist VLAGoogleConsolidation
Generalist VLA
2025
π0.5
Physical Intelligence · Open-world generalization to novel homes
ApproachExtends π0 with heterogeneous training across web data, verbal instruction, and multi-environment robot data — demonstrated cleaning entirely unseen homes it had never encountered during training.
ImpactFirst public demonstration of open-world generalization (not just cross-task, but cross-environment) — a critical milestone toward robots that work in your specific home, not just a demo lab.
Generalist VLAOpen-World
Commercial Deployment
2025
Humanoid Pilot Deployments
Figure, Agility, 1X, Unitree · First paid enterprise pilots
ApproachAgility's Digit begins paid warehouse tote-moving work at logistics customers; Figure signs a multi-year BMW manufacturing deployment; 1X's NEO begins limited home pilot testing.
ImpactMarks the transition from demo videos to revenue-generating deployments — still narrow and heavily supervised, but the first commercial proof points for the entire category.
Commercial DeploymentHumanoid
2026
Architecture
Early 2026
Real-Time On-Device VLA Inference
Industry-wide · Distilled VLA models running fully onboard robots
ApproachDistillation and quantization techniques compress billion-parameter VLA models to run at real-time control frequencies (50–200Hz) directly on robot-onboard compute, removing cloud-latency dependence for safety-critical control loops.
ImpactRemoves network latency and connectivity dependence as a deployment blocker — critical for safety-critical physical actuation in the field.
ArchitectureEfficiency2026
Scale-Up
Mid 2026
First Multi-Thousand-Unit Humanoid Fleets
Industry-wide · Beyond pilot scale into production fleets
ApproachLeading humanoid manufacturers move from dozens of pilot units to fleets in the low thousands across logistics, manufacturing, and early consumer testing — data flywheels from fleet operation begin materially improving model performance.
ImpactThe first real test of whether generalist VLA models hold up at deployment scale outside curated pilot conditions — the industry's next major inflection point.
Scale-Up2026
📊 Why Physical Intelligence, Why Now
$38B
Global humanoid robotics market projection by 2035
$7B+
VC funding into robot foundation model & humanoid startups (2023–2025)
22
Distinct robot embodiments in the Open X-Embodiment cross-training dataset
10M+
US manufacturing/logistics job openings cited as addressable labor gap
Three converging forces explain why "physical intelligence" became a defined field only in the last two years, after decades of comparatively slow robotics progress:
🧠 Transfer Learning Finally Works
Pretrained VLMs already understand "cup," "fragile," "stack" from internet-scale image-text data
Fine-tuning that semantic knowledge onto action requires orders of magnitude less robot-specific data than learning from scratch
Cross-embodiment training (Open X-Embodiment, RT-X) proved skills transfer between different robot bodies
The same transformer architecture and scaling recipes from LLMs now apply directly to action generation
💰 Economic Pull
Persistent labor shortages in warehousing, manufacturing, and eldercare in aging economies
Humanoid form factor reuses existing human-designed infrastructure (doors, stairs, tools) without re-engineering environments
Frontier AI labs (OpenAI, Google) see robotics as the next multi-hundred-billion-dollar market after digital AI
China's manufacturing scale and government subsidy are compressing humanoid hardware costs rapidly
🧩 A Taxonomy of Physical Intelligence
"Robot AI" spans wildly different levels of generality — from a single-task industrial arm to a model that can plan and execute open-ended household chores it was never explicitly trained on. Understanding where a given system sits on this spectrum is the key to evaluating any robotics announcement critically.
By Generality
🔧 Narrow / Task-Specific
Hand-engineered or narrowly-trained for one repeatable task (welding, palletizing, pick-and-place from a fixed bin). Still the vast majority of deployed industrial robots today — reliable, but brittle to any environment change.
🎯 Task-Family Generalist
Generalizes across variations within one task family (any dish arrangement for bussing, any garment for folding) via language-conditioned instruction, but doesn't transfer to unrelated task types without further training.
✨ Open-World Generalist
A single foundation model plans and executes open-ended instructions in novel, unseen environments across many task categories (π0.5, Gemini Robotics, Figure Helix). The frontier target — still limited to relatively structured indoor settings and short-to-medium task horizons in 2026.
By Control Paradigm
🧱 Modular Pipeline
Separate, independently-engineered stages for perception, mapping/planning, and low-level control (classic SLAM-based robotics). Interpretable and debuggable, but errors compound across module boundaries and each stage caps overall generalization.
🌀 End-to-End Learned Policy
A single neural network maps raw sensory input directly to motor commands, trained via imitation learning or reinforcement learning (RT-1, early visuomotor policies). No hand-engineered intermediate representations — but historically limited by scarce training data.
🧬 Vision-Language-Action (VLA)
A pretrained vision-language backbone (with all its internet-scale semantic knowledge) is fine-tuned to output actions instead of just text — the dominant paradigm since RT-2 and the architecture underlying π0, Gemini Robotics, and Figure Helix. Inherits both the strengths and the interpretability challenges of large transformer models.
🗂️ The Model & Robot Landscape
| Model / System |
Company |
Type |
Embodiment |
Control Paradigm |
Access |
| π0 / π0.5 |
Physical Intelligence |
VLA Manipulation |
Cross-embodiment (7+ robots) |
Flow-matching action head |
Research / Enterprise |
| Gemini Robotics |
Google DeepMind |
VLA |
Cross-embodiment |
Native Gemini action head |
Research / Enterprise |
| RT-2 / RT-X |
Google DeepMind |
VLA |
Cross-embodiment (22 robots) |
Action-as-text-tokens |
Research (RT-X open) |
| Figure Helix |
Figure AI |
VLA Humanoid |
Figure 02/03 humanoid |
Dual-system (fast + slow) |
Proprietary / Enterprise |
| NVIDIA GR00T N1 |
NVIDIA |
VLA Humanoid |
Cross-embodiment humanoid |
Dual-system, diffusion action head |
Open Source |
| 1X NEO |
1X Technologies |
Humanoid |
NEO humanoid |
End-to-end learned policy |
Proprietary / Home Pilot |
| Tesla Optimus |
Tesla |
Humanoid |
Optimus Gen 2/3 |
End-to-end (FSD-derived stack) |
Proprietary |
| Boston Dynamics Atlas |
Boston Dynamics (Hyundai) |
Humanoid |
All-electric Atlas |
Model-based + learned control hybrid |
Proprietary / Enterprise |
| Skild Brain |
Skild AI |
VLA |
Embodiment-agnostic |
Foundation model + API |
Enterprise API |
| Unitree G1 / H1 |
Unitree Robotics |
Humanoid Mobile |
G1/H1 humanoid |
Learned locomotion + manipulation |
Commercially available hardware |
| Agility Digit |
Agility Robotics |
Humanoid Mobile |
Digit (bipedal, logistics-focused) |
Task-family generalist |
Enterprise deployment |
| OpenVLA |
Stanford / UC Berkeley / Toyota Research |
VLA |
Cross-embodiment |
Action-as-text-tokens |
Open Source |
⚙️ How VLA Models Actually Work
The dominant architecture — Vision-Language-Action (VLA) — treats robot control as a natural extension of multimodal foundation models: instead of stopping at generating text describing what to do, the model directly generates the motor commands to do it.
📷
Camera + Sensors
RGB, depth, proprioception
🗣️
Language Instruction
"fold the towel"
🧠
Pretrained VLM Backbone
Internet-scale semantic knowledge
🌀
Action Head
Flow-matching / diffusion / tokens
🦾
Motor Commands
Joint torques, 50–200Hz
1. Action Tokenization: Discrete vs. Continuous
Early VLA models (RT-2, OpenVLA) discretized continuous robot actions into text-like tokens — literally reusing the language model's vocabulary to represent joint positions as numbers-as-text. This is elegant and reuses the LLM's existing training machinery, but loses precision and struggles with high-frequency, smooth motion. Newer models (π0, GR00T) instead use a flow-matching or diffusion action head — a small generative model bolted onto the VLM backbone that directly outputs continuous, smooth action sequences (a "chunk" of future actions at once, not one timestep at a time), dramatically improving dexterity and motion quality.
2. Dual-System Architectures (System 1 / System 2)
Inspired by Kahneman's fast/slow thinking framing, several 2024-25 models (Figure Helix, GR00T) split control into two components: a large, slow "System 2" VLM that reasons about the task and environment at ~5-10Hz, and a small, fast "System 1" policy network that translates that reasoning into smooth low-level motor commands at 50-200Hz. This solves a core tension — the semantic reasoning needed for generalization is inherently slower than the control-loop frequency needed for physical stability.
3. Cross-Embodiment Training
Because no single robot generates enough data alone, frontier models train jointly across many different physical robot bodies — different arm lengths, gripper types, camera placements, even bipedal vs. wheeled vs. fixed-base platforms. The model learns embodiment-conditioned representations, allowing skills learned on one robot (e.g., a research arm) to transfer partially to a completely different one (e.g., a humanoid) — the single most important empirical finding enabling the field's recent progress.
🗂️ The Data Problem: Robotics' Core Bottleneck
Unlike text or images, there is no "internet of robot actions" to scrape. Every hour of high-quality robot demonstration data must be physically collected, and this scarcity — not model architecture — is the primary constraint on progress. Labs combine several data sources, roughly in order of cost and fidelity:
Real-World Teleoperation Data
Highest fidelity, most expensive — human operators pilot the actual robot to collect demonstrations
Fleet Deployment Data (autonomous + human correction)
Robots operating in the field, with human oversight correcting failures — scales with deployed fleet size
Simulation + Domain Randomization
Cheap, scalable, infinite variety — but suffers from the sim-to-real gap on contact-rich, fine manipulation tasks
Human Video (egocentric, YouTube, motion capture)
Massive scale, no robot required — but requires solving the "embodiment gap" between human and robot morphology
Internet-Scale Vision-Language Pretraining
No action labels at all — provides the semantic "what is a cup" knowledge the VLA backbone starts from
Cross-embodiment datasets like Open X-Embodiment (1M+ trajectories, 22 robots, 34 institutions) and Physical Intelligence's proprietary multi-robot fleets are the current state of the art for pooling scarce real-world data efficiently. A major open research direction is reducing dependence on expensive teleoperation by better exploiting human video and simulation — the equivalent of finding robotics' own "Common Crawl."
🏢 Major Players
Physical Intelligence
Generalist VLA Foundation Models
π0
π0.5
The category-naming company. Explicitly embodiment-agnostic strategy — builds foundation models, not robots, to be licensed across many hardware partners. $2.4B valuation, backed by OpenAI, Jeff Bezos, Thrive Capital.
Google DeepMind
Research + Native Gemini Integration
RT-2
Gemini Robotics
Open X-Embodiment
Pioneered the VLA paradigm with RT-2 and open-sourced the field's most important cross-embodiment dataset. Now folding robotics directly into the core Gemini multimodal stack — extending consolidation into the physical world.
Figure AI
Vertically Integrated Humanoid
Figure 02/03
Helix VLA
Builds both hardware and the Helix VLA model in-house. Multi-year BMW manufacturing deployment underway. $2.6B valuation with backing from Microsoft, NVIDIA, OpenAI, and Jeff Bezos.
NVIDIA
Infrastructure & Open Foundation Models
GR00T N1
Isaac Lab
Isaac Sim
Positions itself as the compute and tooling layer beneath nearly every humanoid startup — GPUs for training, simulation infrastructure, and an open foundation model any team can fine-tune. Captures value regardless of which robot company wins.
Tesla
Vertically Integrated, FSD-Derived Stack
Optimus Gen 2/3
Leverages Tesla's existing Full Self-Driving neural network infrastructure, data pipeline, and manufacturing scale. Distinctive bet: mass-manufacturing economics from car production applied directly to humanoid hardware costs.
1X Technologies
Consumer/Home-Focused Humanoid
NEO
EVE
Norwegian-founded, OpenAI-backed. Uniquely focused on the home/consumer market rather than industrial deployment first — NEO began limited home pilot testing in 2025, betting on household chores as the proving ground.
Boston Dynamics
Legacy Robotics + Hyundai-Backed
Atlas (electric)
Spot
Decades of best-in-class locomotion and dynamics engineering now being fused with modern learned VLA control. Owned by Hyundai, giving direct access to automotive manufacturing deployment.
Skild AI
Embodiment-Agnostic Foundation Model
Skild Brain
CMU-founded, explicitly pursuing a "foundation model for every robot" strategy similar to Physical Intelligence — licenses its model across quadrupeds, arms, and humanoids rather than building its own hardware. $1.5B valuation.
Unitree Robotics
Low-Cost Hardware Leader (China)
G1
H1
B2 quadruped
Dramatically undercuts Western humanoid pricing (G1 starts under $20K) via Chinese manufacturing scale — commoditizing hardware fast while software/VLA capability still lags US frontier labs.
Agility Robotics
Logistics-First Deployment
Digit
Among the earliest to reach genuine paid commercial deployment — Digit performs tote-moving work at logistics customers today, prioritizing narrow reliability over broad generality as the path to revenue.
🚀 Startups & Funding Activity
Physical intelligence has attracted some of the largest early-stage rounds in AI history — reflecting investor conviction that embodied AI is the next multi-hundred-billion-dollar market, alongside genuine uncertainty about which architectural and go-to-market bets will win.
Physical Intelligence
$400M
Series A · Nov 2024 · Valuation $2.4B
OpenAI, Jeff Bezos, Thrive Capital, Lux Capital
Figure AI
$675M
Series B · Feb 2024 · Valuation $2.6B
Microsoft, NVIDIA, OpenAI, Jeff Bezos, Intel Capital
Skild AI
$300M
Series A · Jul 2024 · Valuation $1.5B
SoftBank, Coatue, Jeff Bezos, Lightspeed
1X Technologies
$100M
Series B · Jan 2024
OpenAI, EQT Ventures, Tiger Global
Agility Robotics
$400M
Series C · 2024 · Valuation ~$1.75B
DCVC, Playground Global, Amazon Industrial Innovation Fund
Apptronik
$403M
Series A · Feb 2025
B Capital, Google, Mercedes-Benz
Dexterity AI
$95M
Series C · 2024 · Valuation ~$1.65B
Lightspeed, Kleiner Perkins
Physical AI (Sanctuary AI)
$140M
Cumulative · 2023–2024
Microsoft, Verizon Ventures
Covariant
$222M
Cumulative · Acquired talent by Amazon 2025
Radical Ventures, Index Ventures, NVIDIA
Investor thesis patterns: the largest checks split between two bets — embodiment-agnostic foundation model companies (Physical Intelligence, Skild AI) that license software across many robot bodies, versus vertically integrated humanoid companies (Figure, Apptronik, Agility) that build hardware and software together to control the full stack. OpenAI has hedged by backing companies on both sides (Physical Intelligence, Figure, and 1X simultaneously). Amazon's acqui-hire of Covariant's team in 2025 signaled that hyperscalers see robotics talent as strategically scarce enough to acquire outright rather than wait for standalone success.
📐 How Physical Intelligence Is Measured
Benchmarking robots is fundamentally harder than benchmarking text or image models — there's no static test set, physical wear affects repeatability, and "success" for a real-world manipulation task is often judged by human evaluators rather than an automated metric.
Real-World Metrics
Task Completion Speed
Figure Helix
MTBF (Reliability)
Agility Digit
Key Metrics Explained
Task Success Rate — Percentage of trials completing the full task correctly, typically human-judged. Frontier VLA models achieve 60–90% on trained task families, dropping sharply on novel objects/environments.
SIMPLER / LIBERO / RoboCasa — Standardized simulation benchmarks approximating real manipulation tasks, letting labs compare models without needing identical physical hardware.
MTBF (Mean Time Between Failures) — Critical for commercial deployment viability, not capability — a robot with 70% task success but frequent hardware failures is less deployable than one with 60% success and high reliability.
Generalization Gap — The performance drop between trained conditions and genuinely novel objects/environments/instructions — the single most-watched metric for judging whether a system is a true generalist or an overfit demo.
🔨 How to Build a Physical Intelligence System From Scratch
Training a frontier generalist VLA model requires access to real robot hardware and significant data-collection infrastructure — well beyond a typical startup's reach for a from-scratch foundation model. But fine-tuning an open VLA (OpenVLA, GR00T) onto a specific task or robot is increasingly tractable, and understanding the full pipeline matters for anyone building in this space.
1. Define Your Generality Target
Decide: narrow task-specific automation (fastest to reliability, least valuable long-term), task-family generalist (fine-tune an open VLA on your domain), or open-world generalist (requires frontier-lab-scale resources). Most startups should fine-tune OpenVLA or GR00T N1 rather than pretrain from scratch — full VLA pretraining costs $10–100M+ in data collection and compute.
2. Choose Your Robot Embodiment
Select hardware matched to your task: fixed-base arms for structured manipulation, mobile manipulators for warehouse/logistics, humanoids for environments designed around human form factors. Consider off-the-shelf platforms (Unitree, commercially available arms) over custom hardware unless embodiment is your core differentiation.
3. Collect Teleoperation Demonstration Data
Set up a teleoperation rig (VR controllers, exoskeleton, or leader-follower arm setup) and collect several hundred to a few thousand demonstrations of your target tasks. Prioritize diversity — varied object positions, lighting, distractor objects — over sheer volume; a diverse 500-demo dataset often outperforms a narrow 5,000-demo one.
4. Supplement with Simulation
Build a simulated version of your task in Isaac Sim, MuJoCo, or similar, and generate large volumes of domain-randomized synthetic demonstrations to supplement scarce real data. Budget for sim-to-real gap — expect meaningfully lower success rates on the physical robot than in simulation until you validate transfer.
5. Fine-tune a Pretrained VLA Backbone
Start from an open-weight VLA (OpenVLA, GR00T N1) rather than training from scratch — this inherits internet-scale semantic understanding your small robot dataset alone could never provide. Fine-tune on your collected demonstrations using LoRA or full fine-tuning depending on how far your task diverges from the base model's training distribution.
6. Choose Your Action Head
For smooth, dexterous manipulation: implement a flow-matching or diffusion action head predicting action "chunks" (10-50 future timesteps at once) rather than single-step actions — dramatically improves motion quality over naive per-timestep prediction.
7. Build a Safety & Constraint Layer
Wrap the learned policy with hard safety constraints — joint limits, force/torque thresholds, collision avoidance, emergency stop triggers — that operate independently of and can override the learned policy. Never rely solely on a learned model for physical safety guarantees.
8. Optimize for Real-Time Control Frequency
Distill or quantize your fine-tuned model to run at the control frequency your task requires (typically 10-50Hz for manipulation, higher for dynamic locomotion). Consider a dual-system split — a fast onboard reactive policy plus a slower cloud-connected planning model — if full on-device inference isn't achievable at your target latency.
9. Evaluate: Success Rate, Generalization Gap, Reliability
Test on held-out object arrangements and, critically, genuinely novel environments not represented in training data — this generalization gap is the most honest signal of real capability versus an overfit demo. Track mechanical reliability (MTBF) alongside task success; a capable-but-fragile system isn't commercially deployable.
10. Build the Data Flywheel for Deployment
Instrument deployed robots to capture failure cases and human corrections, feeding them back into your training pipeline. This deployment data flywheel — not additional pretraining — is typically what closes the gap from "works in the lab" to "works reliably in the field" over 6-18 months of iteration.
💰 Monetization Models
🦾
Robot-as-a-Service (RaaS)
Customers pay a per-hour or per-task rate rather than purchasing hardware outright — mirrors human labor cost structure, lowering the adoption barrier versus capital expenditure.
Agility Digit: reported ~$25/hour warehouse deployment pricing, comparable to a human logistics worker
🏭
Enterprise Deployment Contracts
Multi-year manufacturing/logistics partnerships bundling hardware, model licensing, integration, and support into one enterprise agreement.
Figure × BMW multi-year manufacturing deployment · Boston Dynamics × Hyundai integration
🧠
Foundation Model Licensing
Embodiment-agnostic companies license their VLA model/API to hardware manufacturers rather than building robots themselves — a "software layer" business model.
Skild Brain API, Physical Intelligence model licensing to hardware partners
🛒
Direct Hardware Sales
Traditional capital equipment sale, typically bundled with a software/support subscription for ongoing model updates and fleet management.
Unitree G1: sub-$20K sticker price targeting research and early commercial buyers
🏠
Consumer Subscription (Emerging)
Home robot pilots exploring subscription pricing analogous to a household service, bundling hardware lease with ongoing capability updates.
1X NEO home pilot program — early access subscription structure being tested
⚙️
Infrastructure & Tooling
Compute, simulation, and data infrastructure providers capture value regardless of which robot company ultimately wins the hardware race.
NVIDIA Isaac Sim/Lab licensing, cloud GPU training compute for VLA fine-tuning
Unit economics reality check: a humanoid robot currently costs $30,000–$150,000+ to manufacture depending on dexterity and specification, against a fully-loaded US warehouse worker cost of roughly $35,000–$50,000/year. At current reliability and task-success rates, most deployments still require human supervision or intervention, meaning the honest near-term value proposition is labor augmentation, not full replacement — full ROI parity with human labor at scale is generally projected for 2027-2029 as both hardware costs fall and model reliability improves.
🔬 Open Research Frontiers
🗂️
The Data Scarcity Problem
Without an "internet of robot actions," progress remains bottlenecked by expensive real-world data collection. Better exploitation of human video and simulation — robotics' equivalent of Common Crawl — is the field's most consequential open problem.
🌉
Closing the Sim-to-Real Gap
Contact-rich, fine manipulation tasks (deformable objects, delicate grasping) remain notoriously hard to transfer from simulation to reality. Better physics simulators and learned residual correction models are active research directions.
⏳
Long-Horizon Task Planning
Current generalist models reliably handle tasks spanning seconds to a few minutes; multi-hour tasks requiring memory, error recovery, and hierarchical planning remain largely unsolved.
🖐️
Dexterous Manipulation
Human-level fine manipulation — tying shoelaces, using tools, manipulating cloth — remains substantially harder than locomotion or coarse pick-and-place, and is the primary capability gate for household/caregiving applications.
🛡️
Safety & Liability
Formal safety guarantees for learned policies operating near humans remain an unsolved problem — unlike a chatbot's worst failure mode, a robot's worst failure mode can cause physical harm, demanding fundamentally different verification approaches.
🔋
Power & Actuation Efficiency
Battery life and actuator efficiency limit continuous humanoid operation to hours, not full shifts — a hardware constraint independent of model capability that still gates commercial viability.
🤝
Human-Robot Interaction & Trust
Beyond raw capability, deployment in shared human spaces requires legible, predictable robot behavior that humans can safely anticipate and collaborate with — an under-studied dimension relative to pure task performance.
🌍
World Models for Physical Reasoning
Whether robots need an explicit learned "world model" (predicting physical consequences of actions before executing them) versus purely reactive policies is a live architectural debate connecting directly to the broader world-model research agenda in generative video.
🌐 Applications
📦
Warehousing & Logistics
Tote-moving, palletizing, and order-picking — the current furthest-along commercial deployment segment.
🏭
Manufacturing
Parts assembly, quality inspection, and material handling on production lines, increasingly alongside human workers.
🏠
Home & Domestic
Laundry, dishes, tidying — the hardest and most aspirational target, requiring open-world generalization and fine dexterity.
🏥
Healthcare & Eldercare
Patient mobility assistance, medication delivery, and companionship support amid aging-population labor shortages.
🌾
Agriculture
Selective harvesting, crop monitoring, and precision weeding for tasks too variable for traditional fixed-automation.
☢️
Hazardous Environments
Nuclear facility maintenance, disaster response, and other settings too dangerous for human workers.
🛍️
Retail
Shelf restocking, inventory scanning, and cleaning in retail environments during off-peak hours.
🏗️
Construction
Material transport, site inspection, and repetitive assembly tasks in structured construction environments.
🍽️
Hospitality & Food Service
Table bussing, dishwashing, and kitchen prep support — one of the earliest publicly demonstrated π0 task families.
🔬
Scientific Labs
Repetitive, precision-critical lab procedures — sample handling, pipetting — benefiting from robot consistency over human fatigue.
⚖️ The Verdict: How Far Away Is General Physical Intelligence?
🎯 Our Take
The core scientific bet behind physical intelligence — that internet-pretrained vision-language knowledge transfers into physical action, and that skills generalize across different robot embodiments — is now empirically validated. RT-X, π0, and Gemini Robotics settle that question. This is not the same uncertainty that hung over the field a decade ago.
What remains genuinely uncertain is timeline to reliability at scale. Today's frontier VLA models achieve impressive success rates on trained task families in relatively controlled settings, but the generalization gap to truly novel environments — your specific cluttered kitchen, not a demo lab — is still wide. The data scarcity problem is real and doesn't have an obvious internet-scale shortcut the way text and images did.
The most likely path forward mirrors how autonomous vehicles actually developed: narrow, well-defined commercial deployments first (warehouse logistics, structured manufacturing) where task variability is bounded and ROI is provable today, followed by progressive generalization as deployment data flywheels compound. Open-world home robots — the most publicly exciting target — are likely the last, not first, major application to reach broad reliability, plausibly in the 2028-2030 window rather than 2026-2027.
The strategic read for builders and investors: bet on data flywheel access and narrow deployment ROI now, not raw model capability alone. The company that gets thousands of robots doing real, revenue-generating work in the field — even narrow work — captures the deployment data that compounds into the next generation of capability. That's a very different bet than backing whoever has the most impressive demo video, and it's the one the smartest capital in this space (OpenAI, Bezos, hedging across Physical Intelligence, Figure, and 1X simultaneously) already seems to be making.