Last week we covered the gig workers in Nigeria and India strapping iPhones to their heads to film their chores for $15 an hour. That story went everywhere. But it’s just one slice of a much larger and stranger ecosystem that has quietly emerged around a single problem: how do you teach a robot to exist in the physical world?
The answer, it turns out, involves delivery drivers, skill capture gloves, humans puppeteering robot bodies via VR headsets, and a NVIDIA initiative to simulate physics at planetary scale. Here’s the full picture.
The Problem Every Robotics Company Has to Solve
Teaching a language model to write involves feeding it text. There’s a lot of text on the internet. Teaching a robot to fold a shirt, load a dishwasher, or hand you a cup of coffee is fundamentally different — because the physical world isn’t written down anywhere.
You need motion data. Force data. Spatial data. Video from the robot’s perspective. Joint angles, grip pressures, success rates, failure modes. Hundreds of thousands of examples of humans completing tasks so the robot can build a generalisable model of how physical actions work.
None of that existed in a usable form before the current wave. So the industry invented multiple ways to collect it simultaneously. Some are elegant. Some are unglamorous. All of them are necessary.
1. DoorDash Tasks: 8 Million Couriers, One Data Machine
On March 19, 2026, DoorDash launched a standalone app called Tasks — and quietly repositioned itself from a food delivery company into a data infrastructure provider.
The app pays DoorDash’s 8 million US couriers to complete assignments between deliveries. These aren’t delivery tasks. They’re AI training tasks:
- Film yourself washing at least five dishes, holding each clean dish in frame for a few seconds
- Fold clothes in a natural, unscripted way
- Make a bed, prune plants, repot flowers
- Record yourself having a natural conversation in Spanish with a friend or family member
Pay is shown upfront and varies by effort. Harder tasks — pruning, repotting — pay more. The footage goes to DoorDash’s in-house AI models and those of its partners in retail, insurance, hospitality, and technology.
DoorDash CTO Andy Fang called it plainly: “We think this will be huge for building the frontier of physical intelligence.” (For the full picture on how individual gig workers fit into this, read our piece on Training the Trainers.)
What DoorDash figured out is that the hardest part of collecting physical training data isn’t the technology — it’s the distribution. Getting cameras into diverse homes, environments, and lighting conditions at scale is logistically expensive. DoorDash already solved this for food delivery. They’re now applying that distribution infrastructure to a completely different problem.
Uber did this first. In late 2025, Uber launched a similar initiative allowing US gig workers to earn extra by uploading photos and recordings to train AI models. DoorDash is following the playbook — but with an eight-million-person workforce that could dwarf anything Uber deployed.
2. The Skill Capture Glove: Wearing the Data Collection Device
California-based Sunday Robotics took a different approach. Instead of asking workers to film themselves with a phone, they ship a “skill capture glove” to people across the country.
Workers wear the glove and do household tasks — folding laundry, tidying shelves, cooking, cleaning. The glove records every movement: joint angles, grip pressure, velocity, sequence. When they’re done, they ship it back. Sunday Robotics feeds that motion data into the AI-powered home robot it’s building.
This is more expensive per data point than phone video — but far richer. A phone camera captures what things look like. A skill capture glove captures how actions actually feel and how forces are applied. That’s the data that matters for dexterous manipulation.
Instawork, a staffing app connecting businesses with local hourly workers, has been running a similar programme in Los Angeles — recruiting workers to strap on headbands with phone mounts and clean their homes while recording. The data market is generating an entire hardware cottage industry of wearable capture devices.
3. Teleoperation: Humans Puppeteering Robots in Real Time
The most technically sophisticated approach isn’t cheap consumer hardware — it’s teleoperation: a trained human operator wearing a VR headset and motion capture suit, physically controlling a robot body in real time. Every movement the human makes, the robot mirrors. The robot records everything.
The data quality is exceptional. The cost is high. But companies like Physical Intelligence (π), 1X Technologies, Figure AI, and Apptronik have all built or contracted teleoperation pipelines because some tasks require precision that phone videos simply can’t teach.
The challenge with teleoperation is latency and force feedback. Standard teleoperation systems produce slower, less fluid data because operators can’t feel what the robot is touching. Newer systems are working around this with haptic feedback gloves and improved latency reduction — but it remains the premium tier of physical data collection.
For tasks that require fine manipulation — precision assembly, handling fragile objects — teleoperation data may be the only viable training source.
4. NVIDIA’s Data Factory: Simulating the Physical World at Scale
All of the above methods share a constraint: they’re slow and expensive relative to the amount of data needed. NVIDIA’s answer is to generate data synthetically — at scale — using physics simulation.
NVIDIA’s Open Physical AI Data Factory Blueprint is an open reference architecture that automates how training data is generated, augmented, and evaluated for physical AI systems. The company’s Isaac Sim and Cosmos platforms underpin it, partnering with virtually every major robotics firm — Boston Dynamics, Agility Robotics, and dozens of industrial automation companies.
The fundamental challenge remains the sim-to-real gap: simulations aren’t perfect, and robots trained purely on synthetic data still struggle with unexpected real-world physics. But NVIDIA’s position is that synthetic data can cover 80-90% of training needs, reserving expensive real-world collection for edge cases.
If that thesis holds, NVIDIA doesn’t just sell GPUs to train AI models — it becomes infrastructure for the trillion-dollar AI companies that run on top of it — it sells the simulation infrastructure that makes physical AI training tractable. That’s a platform play worth watching.
5. In-the-Wild Fleet Data: The Robots Teaching Themselves
Once robots are actually deployed at scale, a self-reinforcing data flywheel begins. Every pick, every navigation decision, every failure is logged. The fleet learns from itself continuously.
Tesla is the most ambitious example. With Optimus units operating inside Tesla factories, every unit is simultaneously a worker and a data collector. The factory is the training environment. The robots are the sensors. The product and the research pipeline are the same thing.
This is the end state every robotics company is building toward — but you can only reach it once you’ve cleared the early data hurdle. That’s why the gig economy approaches, teleoperation, and simulation all matter: they’re the bootstrapping mechanisms to get robots capable enough to deploy, at which point the autonomous data loop takes over.
The Emerging Data Economy
What’s taking shape is a multi-tier physical AI data market. No single approach wins. The companies building the best robots in 2027-2028 will be the ones that assembled the best combination — cheap gig data for breadth, teleoperation for precision, simulation for scale, and deployed fleet data for continuous improvement.
The $100 million per year currently flowing through data brokers like Micro1 and Scale AI is a rounding error compared to what this market becomes if humanoid robots reach mass deployment. The real money — in both data collection and data infrastructure — is still ahead.
Why This Matters Beyond Robotics
The physical AI data economy is doing something that hasn’t happened before in tech: it’s creating economic value in developing economies as a direct input to automation in wealthy ones.
A gig worker in Lagos filming her chores for $15/hour is a direct economic participant in the supply chain of robots that may eventually work in Chicago. The extraction flows upward — but unlike most technology supply chains, the labour is geographically distributed and the entry barrier is near-zero.
Whether that’s a feature or a bug depends on your perspective. What’s not in doubt is that the data layer is where physical AI is being built right now — and most people have no idea it exists.
Related Reading
- Training the Trainers: How Gig Workers Around the World Are Teaching Robots to Be Human — Our original deep dive into the phone-on-forehead phenomenon
- Tesla Optimus: The $10 Trillion Bet on Humanoid Robotics — How Tesla is using factory deployment as a training loop
- The Contrarian Case for Perception-First Robotics — Why data collection strategy may matter more than hardware
Sources
- TechCrunch — DoorDash launches Tasks app to pay couriers to train AI
- Bloomberg — DoorDash’s New Paid Tasks Turn Couriers Into AI and Robot Trainers
- Forbes — DoorDash Is Turning 8 Million Couriers Into An AI Training Machine
- NBC News — DoorDash is now letting its drivers train AI on the side
- MIT Technology Review — The gig workers who are training humanoid robots at home
- LA Times — Why people in LA are strapping cameras on their bodies to do chores
- NVIDIA — Global Robotics Leaders Take Physical AI to the Real World
- NVIDIA — Open Physical AI Data Factory Blueprint
- Labellerr — 7 Top Teleoperation Service Providers for Robotics in 2026
- Generalist AI — GEN-1: Scaling Embodied Foundation Models to Mastery
- CNBC — Uber will offer US drivers more gig work including AI data labeling
