Robots don't need more demos. They need more worlds.

Ken Goldberg at UC Berkeley calls robotics' core problem the 100,000-year data gap. That's roughly how long it would take a person to read the text a modern language model trains on. Robots have nothing close.

The instinct is to copy the LLM recipe and collect a mountain of robot data. That race can't be won soon. A possible better plan borrows from what LLM labs did:

1. Piling up robot data won't close the gap.
2. Start from an LLM, then post-train like code.
3. Scale the physical environments.
4. Prove the nines.

1 · Piling up robot data won't close the gap

Robot data is growing fast, and it helps. In April 2026, Generalist trained GEN-1 on more than 500,000 hours of human physical interaction, recorded with low-cost wearable sensors. That's about 57 years of experience, roughly what GPT-2 read in 2019. Open datasets in 2024 and 2025 held a few thousand hours at most. Generalist also found GEN-0's error falls as a power law in data, L(D) = (D_c / D)^α (Generalist, Nov 2025), the same form Kaplan et al. found for language models in 2020. A meta-analysis of 327 papers saw the same across robotics.

Robot training data in 2026 is about where language models were in 2019

Largest training sets in years of experience: hours recorded (robots) or time to read (LLMs).

• Language models: training text
• Robots and wearables: physical interaction

The gap is still enormous. Half a million hours is about 3,600 times less than the text Qwen3 trained on in 2025. JD.com plans to pay 600,000 people to record their movements, aiming for 10 million hours in two years. That's about 1,100 years, still less than GPT-3 read in 2020. The Economist cites an estimate that a robot handling a wide range of daily tasks would need 100 million hours, about Llama 2's reading. HSBC estimates an hour of robot training data in China costs 500 to 700 yuan, or about $70 to $100. At that rate, 100 million hours would cost $7 to $10 billion.

And the power law works against you too. Each halving of error costs a fixed multiple of all the data you already have. More data keeps helping, but every step costs more than the last.

Language models hit this wall first. In December 2024, Ilya Sutskever told NeurIPS that "pre-training as we know it will unquestionably end," because "we have but one internet" (Computing). The labs kept improving anyway by changing what they scaled. Robots have thousands of times less data than LLMs had at that wall, so they should change sooner.

2 · Start from an LLM, then post-train like code

LLMs are worth borrowing from. Physical Intelligence built π0 on Google's PaliGemma to "inherit Internet-scale semantic knowledge," then added more than 10,000 hours of robot data. Google DeepMind built Gemini Robotics on Gemini 2.0. Qwen-RobotManip (June 2026), built on a Qwen vision-language model, ranked first on RoboChallenge's Table30 generalist track.

The knowledge transfers. π0.7 (April 2026) ran an air fryer it had seen in only two training episodes, coached step by step in plain language. Frontier models can even drive a robot with no robot training. In a September 2026 study, OpenAI's GPT-6 Astra, used directly as the policy, averaged 22.5% on the RoboDojo benchmark over 2,100 simulated trials, beating every public entry, including π0.5 at 6.9%. Stanford's HomeBody used Astra to direct a Unitree G1 humanoid as it tidied a kitchen it had never seen. Claude Opus 4.7 programmed a robot dog through a set of tasks about 20 times faster than the fastest human team in 2025.

But borrowing is not enough. A robot that works 60–70% of the time is a demo. It still needs someone standing by. Andrej Karpathy calls the road from demo to product a "march of nines": 90% is "just the first nine," and "every single nine is a constant amount of work." In five years at Tesla, he said, self-driving got through "maybe three nines or two nines" (Electrek).

Most robot learning is still on the first nine. On RoboChallenge's Table30, 30 everyday tasks on real robots, the best model at launch in October 2025 averaged 43.7% when fine-tuned per task. A generalist π0.5 managed 17.7%. Frontier LLMs fall short on real hardware too. Astra struggled with precise and two-handed tasks, and its real-robot tests were stopped after it "repeatedly issued physically unreasonable or unsafe actions." Anthropic's own tests found Claude far better at high-level control than low-level, and estimated real-time control needs about 83 commands a second, against 0.2 to 0.4 for frontier models. The LLM knows what to do. It doesn't yet know how.

Some of that can come from humans instead of robots. GEN-1 pretrained on wearable data alone and then needed about an hour of robot data per task. The rest has to come from practice.

Post-training is the answer. Language models have been through this. A pretrained model knows a lot but can't reliably do anything with it. Ask it to fix a bug and it writes plausible code that doesn't run. What made it useful was practice: hand it a repo and a broken test, let it try, check the result, and do that over and over.

Robots learn the same way. With Recap, Physical Intelligence let π*0.6 practice on its own, learning from mistakes and from an expert stepping in when it got stuck. On some of the hardest tasks it more than doubled throughput and failed half as often. It once made espresso from 5:30 in the morning to 11:30 at night. Berkeley's HIL-SERL reached near-perfect success on real tasks in one to two and a half hours of practice. A single π0.7 model now matches those specialists on laundry, espresso and boxes.

So for robots, the thing to scale isn't data you read once. It's places to practice. A coding model can practice in a million copies of the same repo at once. A robot brain needs a body, and a real kitchen with real cups to drop.

3 · Scale the physical environments

More environments means more rooms, objects, robots, and a way to reset them all. That's the bottleneck. It's also the lever.

Lin et al. (ICLR 2025) collected more than 40,000 demonstrations and ran 15,000 real rollouts to see what drives generalization. Error fell as a power law in the number of training environments. More demos in the same setting plateaued. Each doubling of environments cut error 28–44%. (Each doubling of tokens cuts LLM loss about 6%, though the units aren't comparable.) With about 50 demos in each of 32 environment-object pairs, their policies succeeded 85–92.5% in rooms and with objects they'd never seen. Robots should count environments the way LLM labs count tokens.

Robot error falls with environments the way LLM loss falls with tokens

Both are power laws: straight lines on log-log axes. Dots are measured; lines are fitted. In these fits, each doubling of environments cuts robot error 28–44%, and each doubling of tokens cuts LLM loss about 6%.

Language models

Test loss (nats per token) · Kaplan et al. 2020

• Fitted law · loss ∝ D^−0.095

Robot policies

Error (1 − normalized score) · Lin et al., ICLR 2025

• Pour water ∝ E^−0.844
• Mouse arrangement ∝ E^−0.466

Physical Intelligence saw the same at scale. π0.5's success in unseen homes rose with every added environment and nearly matched in-home training at about 100. Removing its cross-environment data dropped out-of-distribution success from 94% to 31%. π0.7, trained on many robots, human recordings and autonomous runs, folded laundry on a robot that had never had laundry data, matching expert teleoperators using that robot for the first time.

Why not only simulation?

Coding labs spin up containers, so why not spin up simulated kitchens? Because a container running a repo is the software. A simulated kitchen is a guess at a kitchen.

Simulation works best where the physics is simple. Goldberg notes it has worked for locomotion and acrobatics but stayed "elusive" for dexterity (Berkeley News). A survey of real-world deep RL by Tang et al. finds the same split: zero-shot sim-to-real is the most mature recipe for legged locomotion, while open-world and contact-rich manipulation lags, partly because simulators handle contact poorly. Even a famous sim-trained success shows the ceiling. OpenAI's robot hand, trained entirely in simulation, solved a Rubik's Cube about 60% of the time, and about 20% on the hardest scrambles.

And a Rubik's Cube is a rigid block. Kitchens are full of things that bend, pour, tear and squish. Cloth, rope and cables have effectively infinite degrees of freedom and change shape as you touch them (McKenna and Oyekan, 2026; SoftGym). A 2024 benchmark of MuJoCo, Bullet, Flex and SOFA on cloth noted that few works had even measured how far these simulators drift from real fabric. Most robot fluid work assumes an idealized fluid in simple pouring tasks, far from the mixing, sticking and splashing of a real kitchen (FluidLab). Cutting a tomato means simulating a blade, soft material and a crack spreading through it at once. NVIDIA's DiSECt built a dedicated simulator just for that. Other senses are missing too: sound, which tells you a glass is nearly full, is "notoriously difficult to simulate," and sim-to-real has mostly worked for vision alone (Wang et al., 2025).

The weak spot is physics, not pictures. A June 2026 position paper from Stanford, ETH Zurich, TU Darmstadt and UCL put it this way: "A visually plausible rollout that ignores contact, force, friction, or stability is not yet reliable robot supervision."

So the simulators that work are tuned against the real world. DiSECt only matched reality after it was calibrated to forces and deformations from real cutting data. RialTo builds its training sims by scanning real rooms and importing real demos, and reports over 67% gains in robustness from that real-to-sim-to-real loop. Evaluation works the same way. SIMPLER matched real-world policy rankings closely (Pearson r around 0.9) only after its authors closed the control and visual gaps to the real setups. WorldEval's video model was fine-tuned on 1,400 real trajectories and checked against more than 1,000 real trials. Google DeepMind's Veo-based simulator for Gemini Robotics was validated with more than 1,600 real evaluations, and GigaWorld-1 against real rollouts.

Sim is good for cheap exploration. Real rooms are where you calibrate it, find what it missed, and prove a policy works.

4 · Prove the 9s

Practice and breadth push a robot up the nines. GEN-1 averaged 99% on tasks where earlier models managed 64%, which is two nines on Generalist's own tasks. But a success rate is only as good as the trials behind it.

Consecutive successes needed to claim that rate with 95% confidence, by the rule of three (about 3 ÷ failure rate).

Success rate	Failures at 100 tasks a day	Clean runs in a row to show it
70%	30 a day	about 10
90%	10 a day	about 30
99%	1 a day	about 300
99.9%	1 every 10 days	about 3,000
99.99%	1 every 100 days	about 30,000

Each nine takes ten times more trials to prove. A perfect 10 out of 10 only shows a true rate above about 74%. Proving 99.99% takes about 30,000 clean runs, or 1,000 robot-hours at two minutes a run, for one task, and every release resets the count. The failures left are also the rare ones: an odd object, glare on a counter, a sticky drawer. You only find them by running in many places, which brings us back to step 3.

Coding models are graded on held-out tests every release. Robots need the same, but their tests are physical. Earning the nines and proving them both take real robot time in many real places. Right now only teams with their own facilities have that.

Environments as infrastructure

Most teams don't build a data center to train a model. They rent an instance.

2d9s does that for the physical world: thousands of real robots in instrumented environments, shared across teams and rented by the hour. Configure a scene, launch one robot or a fleet through an API, and teleoperate from a browser or push a policy straight to the hardware.

• Post-training loops. Deploy, practice, collect failures and corrections, retrain, redeploy. Nightly instead of quarterly.
• Breadth on demand. Demos, rollouts, corrections, and force and contact data across many environments at once.
• Evaluation that counts nines. Standard task suites, neutral scorecards and regression tests in environments you didn't train in. RoboChallenge had already run more than 40,000 remote real-robot tests by April 2026.
• No wasted hours. The Economist reports eight hours of robot coffee-pouring yields about three useful ones. A shared fleet stays busy, and failures become training data.
• Sim and real together. A mirrored simulation for cheap exploration, real rigs to keep it honest.
• Your hardware, our worlds. Bring your own robot and run it as a private instance type.

What this means for your team

1. Spend robot time on post-training. Real-world practice is the scarce resource.
2. Build environments like test suites. A scene, a success check, a reset. Sim to explore, real rooms to decide what ships.
3. Count environments, not hours. A few dozen demos in many places beat hundreds in one.
4. Report nines where you didn't train. Hold out real environments and test every release.
5. Rent breadth instead of building it. Your edge is your model, not your floor space.

Robots won't read their way across the 100,000-year gap. They'll borrow what language models know, then practice in thousands of real places until failure is rare. That only works if every team can get to those places.