Figure AI - Helix 2.5 Robot: Zero-Shot Home Generalization
5 hours ago
- Helix 2.5 is the most advanced neural network neural network developed by Figure, focusing on whole-body autonomy in humanoid robots.
- The model achieves zero-shot generalization, allowing robots to perform tasks in unfamiliar homes without prior data collection or adaptation.
- Index pretraining, trained on global-scale human behavior data, was key to Helix 2.5's 56% zero-shot success rate, compared to 9% from scratch.
- A single foundation model supports three behaviors: tidying living rooms, folding towels, and making beds across 30 unseen homes.
- Helix 2.5 requires half as much task-specific data as Helix 02, yet achieves comparable or better generalization across a 30-fold broader scope.
- The research introduces a human-to-humanoid robot transfer scaling law, showing smooth improvements in downstream action prediction with more Index pretraining data.
- The model exhibits improved self-correction abilities, enabling it to recover from errors like misaligned folds during long tasks.
- Success criteria for evaluation are strict, requiring full task completion with detailed rubrics for scoring pillows, comforters, and towels.
- The findings suggest that scaling up data and compute in pretraining can lead to more physical world understanding for robots, supporting the thesis that robotic generalization is achievable.