Figure released the humanoid robot neural network Helix2.5 on September 17. The model was pre-trained based on the company's Index human behavior dataset and tested in 30 households in the San Francisco Bay Area that had not been previously collected. The results showed that compared to strategies trained from scratch under the same conditions, Index pre-training increased the zero-shot success rate from 9% to 56%.
Helix2.5 is based on a single Index pre-trained base model and supports three full-body actions: organizing the living room, folding towels, and making the bed. During the evaluation, the home environment and toys, towels, and bedding were not included in the task specification data, and all tasks required complete completion to be considered successful. Figure also stated that Helix2.5 achieved comparable success rates using only half the adaptive data of the previous Helix02, and demonstrated full-body self-correction capabilities such as retreating to adjust position, changing stance, and moving on the bed.
To isolate the contribution of pre-training, Figure kept the model architecture, optimization method, hyperparameters, downstream data, and evaluation conditions consistent, only changing the initialization method. The company claims that no single evaluation task in the Index dataset accounted for more than 1.90%, and the above results directly reflect the contribution of pre-training to generalization ability. Further experiments showed that when the amount of Index data doubled, the action prediction loss showed a predictable downward trend, and the company used this to measure the data scaling law in the humanoid robot field.
Index was officially launched on August 25, with a total of 264,000 downloads so far, over 44,000 weekly active users, and has collected over 16 million videos, paying $15 million to data contributors. Figure said it has invested $3.5 billion in computing resources to train the Helix series.
