Figure AI Claims Robot Scaling Law; Peers Say 56% Success Proves No Generalization

Figure AI released Helix 2.5, which founder Brett Adcock calls the company's most important project since its founding. With a $39 billion valuation, Figure AI is the world's most valuable private humanoid robotics company.

Past Helix demos drew criticism for taking place in well-lit model homes. Helix 2.5 aims to move beyond that by renting 30 residences in the San Francisco Bay Area and shipping robots directly to the sites.

The robots were tasked with tidying living rooms, making beds, and folding towels. The team had not previously collected robot data in these homes or fine-tuned on the beds, pillows, towels, and toys found there.

Each of the three tasks was tested 140 times. Tidying the living room required putting all 13 to 15 toys into a basket; folding towels required neatly folding four towels and placing them in a basket; making the bed required meeting pillow position, blanket direction, and surface coverage requirements. Any need for human safety intervention or exceeding the time limit was recorded as a failure.

Bed-making succeeded 94 times (about 67% success); towel folding succeeded 87 times (about 62%); toy tidying succeeded 56 times (40%). Combined, 237 completions yielded an overall task success rate of 56%. In the same tests, a control model without Index pretraining succeeded only about 9% of the time.

Figure AI trained models on four Index dataset sizes (1x, 2x, 4x, and 8x), fixing model size, downstream task data, and evaluation method while only increasing human behavior pretraining data. Each doubling of data continued to reduce the model's error in predicting the robot's next action.

Using the first three experiments to predict the largest-scale training, the final error was only 0.54% of the entire range of variation. Figure AI calls it the first measured human-to-robot transfer scaling law on humanoid robots. The company has signed a compute agreement with Nscale including a $3.5 billion initial commitment, with deployment planned from the second half of 2027 and a long-term goal of up to 100,000 Nvidia Vera Rubin GPUs. Index continues to collect human experience at about 35 minutes per second.

Peer criticism: Tony Zhao, co-founder of Sunday Robotics and a core author of the ACT and ALOHA systems, focused on the 56% reliability. After Figure AI released the data, he wrote that useful work equals generalization plus reliability. Figure AI interprets the 237 successes as the robot being able to work in unfamiliar environments; Tony focuses on the other 183 failures.

In home environments, robots may encounter fragile items and may interact with children and pets. Whether a robot can occasionally complete a task and whether users can trust it to work independently are two different things. Nikolai Ensslen, co-founder and former CEO of robot motion control company Synapticon, says the numbers themselves are evidence that imitation learning cannot generalize. In his definition, generalization means the system still works in a different place and works every time.

The controversy is whether this transfer is sufficient to support Figure AI's judgment about the future. If 56% can steadily rise to 80%, 90%, or higher as Index scales, Helix 2.5 will become a starting point for the robot pretraining route. If the success rate remains near half, then the criticism from Tony Zhao and Nikolai Ensslen will become more than an impolite comment—it will question the entire route related to imitation learning.