Figure Announces Humanoid Helix 2.5

Figure has announced Helix 2.5, a humanoid model pre-trained on the human behavior dataset 'Index'. This model demonstrated zero-shot generalization capabilities by performing household tasks such as tidying toys, folding towels, and making beds in 30 homes in the Bay Area that were not included in the training data.

Pre-training Effect

Comparing performance differences based on pre-training under identical task data and evaluation conditions, the Index pre-trained model showed higher performance compared to the randomly initialized model.

• Folding Towels: Success rate 9% → 62%
• Tidying Toys: Success rate 5% → 40%
• Making Beds: Success rate 11% → 67%
• Aggregate Success Rate: Approx. 8% → 56%

This demonstrates that pre-training on broad human behavior data contributes to improved robot manipulation performance in new environments. (The original text describes an increase from 9% to 56%, while the chart calculation values are approximately 8.3% to 56.4%)

Technical Features and Limitations

Helix 2.5 handles complex tasks requiring Locomanipulation combining locomotion and manipulation, as well as active perception. Figure explains that, unlike Helix 02, it achieved similar success rates with half the amount of task adaptation data without using data from the evaluation environment. Additionally, they confirmed a Scaling Law where Action-Prediction Loss decreases in a predictable pattern as the volume of human behavior data increases.

However, since the evaluation was limited to a specific region (Bay Area) and three tasks, further research is needed to verify overall generalization. Model weights and reproduction code have not been released.