The 100-billion-hour experience gap: where robot data gets stuck

For physical AI to become truly autonomous, one prerequisite is roughly 100 billion hours of validated real-world experience. Betta gives this order of magnitude in a tweet and also says only 0.01% has been completed so far. A quoted passage in the same material says 0.1% instead. The two numbers differ by 10x; set that aside for now—it does not affect the conclusion: we are still far from the goal.

What Inverted Lambda is doing is one layer removed from “making robots smart on their own.” It is a decentralized teleoperation platform built on Base. Users operate robots remotely, complete episodes one by one, and earn points. The team opens access to the public and distributes a 4,000 USDC reward pool to participants.

The reward pool is split into four parts: Robot Operations 1,900 USDC for the top 15; Shaping the Vision 1,050 USDC for the top 15; Referral Network 650 USDC for the top 10; Lucky Draw 400 USDC for 20 winners.

The third phase is called Third Contact and runs from September 26 to November 6, 2026. Betta says the focus of phase three is hyper-realistic simulation. The more an environment resembles the real world, the more grounded operators’ tactile judgments are when handling objects such as grippers, plush toys, and faucets.

Phase three adds Evaluation. Participants do more than teleoperation; they also score teleoperation recordings. The scores determine which data qualifies for training models. Each user can evaluate at most 100 videos per day.

Scoring uses ABCD. A means the task was fully completed with no pauses, retries, or drops. B means completed, but with 1 to 2 retries, or cumulative pauses, task deviations, or going out of frame of no more than 10 seconds. C means partially completed, with 3 or more drops, retries, or cumulative pauses, task deviations, or going out of frame exceeding 10 seconds. D means complete failure, unusable, or the recording cannot be used.

This design addresses a specific problem: not every completed operation can serve as training data. Data for physical AI is slow and expensive, and the material mentions three product updates all centered on this. real2sim brings real-world information into simulation so developers do not have to repeat every experiment. Physical Agent SDK aims to make it easier for developers to build physical AI experiments. Data Pipeline replays, retargets, and scores all episodes to filter out clean, usable data.

If you want to participate, the process is not complicated. Visit inverted-lambda.com and enter a referral code, connect X and Discord, complete social tasks, then choose easy, normal, or hard based on experience. Each task has 20 runs. Teleoperation does not require a subscription.

In the summary, Betta writes that the first contact brings people into the experience, the second uses real objects to push limits, and the third hands judgment over to users. Collecting thousands of robot episodes means very little if the quality does not meet the bar. This judgment is more worth watching than the reward pool.

Third Contact ends on November 6, 2026. The experience-sharing in the material comes from Betta’s tweet, and the rewards and rules are also subject to that tweet.